Paolo Soda

Irene Iele, Giulia Romoli, Daniele Molino et al.

Short-term forecasting of vegetation dynamics is a key enabler for data-driven decision support in precision agriculture. Normalized Difference Vegetation Index (NDVI) forecasting from satellite observations, however, remains challenging due to sparse and irregular sampling caused by cloud masking, as well as the heterogeneous climatic conditions under which crops evolve. In this work, we propose a probabilistic forecasting framework for field-level NDVI prediction under sparse, irregular clear-sky acquisitions. The architecture separates the encoding of historical NDVI and meteorological observations from future exogenous covariates, fusing both representations for multi-step quantile prediction. To address irregular revisit patterns and horizon-dependent uncertainty, we introduce a temporal-distance weighted quantile loss that aligns the training objective with the effective forecasting horizon. In addition, we incorporate cumulative and extreme-weather feature engineering to capture delayed meteorological effects relevant to vegetation response. Experiments on European satellite data show that the proposed approach outperforms statistical, deep learning, and time-series baselines on both pointwise and probabilistic evaluation metrics. Ablation studies confirm that target history is the primary driver of performance, with meteorological covariates providing additional gains in the full multimodal setting. The code is available at https://github.com/arco-group/ndvi-forecasting.

Camillo Maria Caruso, Paolo Soda, Valerio Guarrasi

Handling missing values in tabular datasets presents a significant challenge in training and testing artificial intelligence models, an issue usually addressed using imputation techniques. Here we introduce "Not Another Imputation Method" (NAIM), a novel transformer-based model specifically designed to address this issue without the need for traditional imputation techniques. NAIM's ability to avoid the necessity of imputing missing values and to effectively learn from available data relies on two main techniques: the use of feature-specific embeddings to encode both categorical and numerical features also handling missing inputs; the modification of the masked self-attention mechanism to completely mask out the contributions of missing data. Additionally, a novel regularization technique is introduced to enhance the model's generalization capability from incomplete data. We extensively evaluated NAIM on 5 publicly available tabular datasets, demonstrating its superior performance over 6 state-of-the-art machine learning models and 5 deep learning models, each paired with 3 different imputation techniques when necessary. The results highlight the efficacy of NAIM in improving predictive performance and resilience in the presence of missing data. To facilitate further research and practical application in handling missing data without traditional imputation methods, we made the code for NAIM available at https://github.com/cosbidev/NAIM.

Lorenzo Tronchin, Minh H. Vu, Paolo Soda et al.

Data Augmentation (DA) is a technique to increase the quantity and diversity of the training data, and by that alleviate overfitting and improve generalisation. However, standard DA produces synthetic data for augmentation with limited diversity. Generative Adversarial Networks (GANs) may unlock additional information in a dataset by generating synthetic samples having the appearance of real images. However, these models struggle to simultaneously address three key requirements: fidelity and high-quality samples; diversity and mode coverage; and fast sampling. Indeed, GANs generate high-quality samples rapidly, but have poor mode coverage, limiting their adoption in DA applications. We propose LatentAugment, a DA strategy that overcomes the low diversity of GANs, opening up for use in DA applications. Without external supervision, LatentAugment modifies latent vectors and moves them into latent space regions to maximise the synthetic images' diversity and fidelity. It is also agnostic to the dataset and the downstream task. A wide set of experiments shows that LatentAugment improves the generalisation of a deep model translating from MRI-to-CT beating both standard DA as well GAN-based sampling. Moreover, still in comparison with GAN-based sampling, LatentAugment synthetic samples show superior mode coverage and diversity. Code is available at: https://github.com/ltronchin/LatentAugment.

Alessandro Pesci, Valerio Guarrasi, Marco Alì et al.

The translation from Magnetic resonance imaging (MRI) to Computed tomography (CT) has been proposed as an effective solution to facilitate MRI-only clinical workflows while limiting exposure to ionizing radiation. Although numerous Generative Adversarial Network (GAN) architectures have been proposed for MRI-to-CT translation, systematic and fair comparisons across heterogeneous models remain limited. We present a comprehensive benchmark of ten GAN architectures evaluated on the SynthRAD2025 dataset across three anatomical districts (abdomen, thorax, head-and-neck). All models were trained under a unified validation protocol with identical preprocessing and optimization settings. Performance was assessed using complementary metrics capturing voxel-wise accuracy, structural fidelity, perceptual quality, and distribution-level realism, alongside an analysis of computational complexity. Supervised Paired models consistently outperformed Unpaired approaches, confirming the importance of voxel-wise supervision. Pix2Pix achieved the most balanced performance across districts while maintaining a favorable quality-to-complexity trade-off. Multi-district training improved structural robustness, whereas intra-district training maximized voxel-wise fidelity. This benchmark provides quantitative and computational guidance for model selection in MRI-only radiotherapy workflows and establishes a reproducible framework for future comparative studies. To ensure the reproducibility of our experiments we make our code public, together with the overall results, at the following link:https://github.com/arco-group/MRI_TO_CT.git

Valerio Guarrasi, Fatih Aksu, Camillo Maria Caruso et al.

Deep learning has revolutionized biomedical research by providing sophisticated methods to handle complex, high-dimensional data. Multimodal deep learning (MDL) further enhances this capability by integrating diverse data types such as imaging, textual data, and genetic information, leading to more robust and accurate predictive models. In MDL, differently from early and late fusion methods, intermediate fusion stands out for its ability to effectively combine modality-specific features during the learning process. This systematic review aims to comprehensively analyze and formalize current intermediate fusion methods in biomedical applications. We investigate the techniques employed, the challenges faced, and potential future directions for advancing intermediate fusion methods. Additionally, we introduce a structured notation to enhance the understanding and application of these methods beyond the biomedical domain. Our findings are intended to support researchers, healthcare professionals, and the broader deep learning community in developing more sophisticated and insightful multimodal models. Through this review, we aim to provide a foundational framework for future research and practical applications in the dynamic field of MDL.

Matteo Tortora, Ermanno Cordelli, Rosa Sicilia et al.

The current cancer treatment practice collects multimodal data, such as radiology images, histopathology slides, genomics and clinical data. The importance of these data sources taken individually has fostered the recent raise of radiomics and pathomics, i.e. the extraction of quantitative features from radiology and histopathology images routinely collected to predict clinical outcomes or to guide clinical decisions using artificial intelligence algorithms. Nevertheless, how to combine them into a single multimodal framework is still an open issue. In this work we therefore develop a multimodal late fusion approach that combines hand-crafted features computed from radiomics, pathomics and clinical data to predict radiation therapy treatment outcomes for non-small-cell lung cancer patients. Within this context, we investigate eight different late fusion rules (i.e. product, maximum, minimum, mean, decision template, Dempster-Shafer, majority voting, and confidence rule) and two patient-wise aggregation rules leveraging the richness of information given by computer tomography images and whole-slide scans. The experiments in leave-one-patient-out cross-validation on an in-house cohort of 33 patients show that the proposed multimodal paradigm with an AUC equal to $90.9\%$ outperforms each unimodal approach, suggesting that data integration can advance precision medicine. As a further contribution, we also compare the hand-crafted representations with features automatically computed by deep networks, and the late fusion paradigm with early fusion, another popular multimodal approach. In both cases, the experiments show that the proposed multimodal approach provides the best results.

Aurora Rofena, Valerio Guarrasi, Marina Sarli et al.

Contrast Enhanced Spectral Mammography (CESM) is a dual-energy mammographic imaging technique that first needs intravenously administration of an iodinated contrast medium; then, it collects both a low-energy image, comparable to standard mammography, and a high-energy image. The two scans are then combined to get a recombined image showing contrast enhancement. Despite CESM diagnostic advantages for breast cancer diagnosis, the use of contrast medium can cause side effects, and CESM also beams patients with a higher radiation dose compared to standard mammography. To address these limitations this work proposes to use deep generative models for virtual contrast enhancement on CESM, aiming to make the CESM contrast-free as well as to reduce the radiation dose. Our deep networks, consisting of an autoencoder and two Generative Adversarial Networks, the Pix2Pix, and the CycleGAN, generate synthetic recombined images solely from low-energy images. We perform an extensive quantitative and qualitative analysis of the model's performance, also exploiting radiologists' assessments, on a novel CESM dataset that includes 1138 images that, as a further contribution of this work, we make publicly available. The results show that CycleGAN is the most promising deep network to generate synthetic recombined images, highlighting the potential of artificial intelligence techniques for virtual contrast enhancement in this field.

Chongsheng Zhang, George Almpanidis, Gaojuan Fan et al.

Long-tailed data is a special type of multi-class imbalanced data with a very large amount of minority/tail classes that have a very significant combined influence. Long-tailed learning aims to build high-performance models on datasets with long-tailed distributions, which can identify all the classes with high accuracy, in particular the minority/tail classes. It is a cutting-edge research direction that has attracted a remarkable amount of research effort in the past few years. In this paper, we present a comprehensive survey of latest advances in long-tailed visual learning. We first propose a new taxonomy for long-tailed learning, which consists of eight different dimensions, including data balancing, neural architecture, feature enrichment, logits adjustment, loss function, bells and whistles, network optimization, and post hoc processing techniques. Based on our proposed taxonomy, we present a systematic review of long-tailed learning methods, discussing their commonalities and alignable differences. We also analyze the differences between imbalance learning and long-tailed learning approaches. Finally, we discuss prospects and future directions in this field.

Valerio Guarrasi, Lorenzo Tronchin, Domenico Albano et al.

We are witnessing a widespread adoption of artificial intelligence in healthcare. However, most of the advancements in deep learning in this area consider only unimodal data, neglecting other modalities. Their multimodal interpretation necessary for supporting diagnosis, prognosis and treatment decisions. In this work we present a deep architecture, which jointly learns modality reconstructions and sample classifications using tabular and imaging data. The explanation of the decision taken is computed by applying a latent shift that, simulates a counterfactual prediction revealing the features of each modality that contribute the most to the decision and a quantitative score indicating the modality importance. We validate our approach in the context of COVID-19 pandemic using the AIforCOVID dataset, which contains multimodal data for the early identification of patients at risk of severe outcome. The results show that the proposed method provides meaningful explanations without degrading the classification performance.

Valerio Guarrasi, Paolo Soda

The COVID-19 pandemic has caused millions of cases and deaths and the AI-related scientific community, after being involved with detecting COVID-19 signs in medical images, has been now directing the efforts towards the development of methods that can predict the progression of the disease. This task is multimodal by its very nature and, recently, baseline results achieved on the publicly available AIforCOVID dataset have shown that chest X-ray scans and clinical information are useful to identify patients at risk of severe outcomes. While deep learning has shown superior performance in several medical fields, in most of the cases it considers unimodal data only. In this respect, when, which and how to fuse the different modalities is an open challenge in multimodal deep learning. To cope with these three questions here we present a novel approach optimizing the setup of a multimodal end-to-end model. It exploits Pareto multi-objective optimization working with a performance metric and the diversity score of multiple candidate unimodal neural networks to be fused. We test our method on the AIforCOVID dataset, attaining state-of-the-art results, not only outperforming the baseline performance but also being robust to external validation. Moreover, exploiting XAI algorithms we figure out a hierarchy among the modalities and we extract the features' intra-modality importance, enriching the trust on the predictions made by the model.

Camillo Maria Caruso, Valerio Guarrasi, Sara Ramella et al.

In the field of lung cancer research, particularly in the analysis of overall survival (OS), artificial intelligence (AI) serves crucial roles with specific aims. Given the prevalent issue of missing data in the medical domain, our primary objective is to develop an AI model capable of dynamically handling this missing data. Additionally, we aim to leverage all accessible data, effectively analyzing both uncensored patients who have experienced the event of interest and censored patients who have not, by embedding a specialized technique within our AI model, not commonly utilized in other AI tasks. Through the realization of these objectives, our model aims to provide precise OS predictions for non-small cell lung cancer (NSCLC) patients, thus overcoming these significant challenges. We present a novel approach to survival analysis with missing values in the context of NSCLC, which exploits the strengths of the transformer architecture to account only for available features without requiring any imputation strategy. More specifically, this model tailors the transformer architecture to tabular data by adapting its feature embedding and masked self-attention to mask missing data and fully exploit the available ones. By making use of ad-hoc designed losses for OS, it is able to account for both censored and uncensored patients, as well as changes in risks over time. We compared our method with state-of-the-art models for survival analysis coupled with different imputation strategies. We evaluated the results obtained over a period of 6 years using different time granularities obtaining a Ct-index, a time-dependent variant of the C-index, of 71.97, 77.58 and 80.72 for time units of 1 month, 1 year and 2 years, respectively, outperforming all state-of-the-art methods regardless of the imputation method used.

Matteo Tortora, Francesco Conte, Gianluca Natrella et al.

Accurate forecasting of renewable generation is crucial to facilitate the integration of RES into the power system. Focusing on PV units, forecasting methods can be divided into two main categories: physics-based and data-based strategies, with AI-based models providing state-of-the-art performance. However, while these AI-based models can capture complex patterns and relationships in the data, they ignore the underlying physical prior knowledge of the phenomenon. Therefore, in this paper we propose MATNet, a novel self-attention transformer-based architecture for multivariate multi-step day-ahead PV power generation forecasting. It consists of a hybrid approach that combines the AI paradigm with the prior physical knowledge of PV power generation of physics-based methods. The model is fed with historical PV data and historical and forecast weather data through a multi-level joint fusion approach. The effectiveness of the proposed model is evaluated using the Ausgrid benchmark dataset with different regression performance metrics. The results show that our proposed architecture significantly outperforms the current state-of-the-art methods. These findings demonstrate the potential of MATNet in improving forecasting accuracy and suggest that it could be a promising solution to facilitate the integration of PV energy into the power grid.

Francesco Di Feola, Lorenzo Tronchin, Paolo Soda

The current deep learning approaches for low-dose CT denoising can be divided into paired and unpaired methods. The former involves the use of well-paired datasets, whilst the latter relaxes this constraint. The large availability of unpaired datasets has raised the interest in deepening unpaired denoising strategies that, in turn, need for robust evaluation techniques going beyond the qualitative evaluation. To this end, we can use quantitative image quality assessment scores that we divided into two categories, i.e., paired and unpaired measures. However, the interpretation of unpaired metrics is not straightforward, also because the consistency with paired metrics has not been fully investigated. To cope with this limitation, in this work we consider 15 paired and unpaired scores, which we applied to assess the performance of low-dose CT denoising. We perform an in-depth statistical analysis that not only studies the correlation between paired and unpaired metrics but also within each category. This brings out useful guidelines that can help researchers and practitioners select the right measure for their applications.

Elena Mulero Ayllón, Linlin Shen, Pierangelo Veltri et al.

Accurate lung tumor segmentation is vital for improving diagnosis and treatment planning, and effectively combining anatomical and functional information from PET and CT remains a major challenge. In this study, we propose vMambaX, a lightweight multimodal framework integrating PET and CT scan images through a Context-Gated Cross-Modal Perception Module (CGM). Built on the Visual Mamba architecture, vMambaX adaptively enhances inter-modality feature interaction, emphasizing informative regions while suppressing noise. Evaluated on the PCLT20K dataset, the model outperforms baseline models while maintaining lower computational complexity. These results highlight the effectiveness of adaptive cross-modal gating for multimodal tumor segmentation and demonstrate the potential of vMambaX as an efficient and scalable framework for advanced lung cancer analysis. The code is available at https://github.com/arco-group/vMambaX.

Alice Natalina Caragliano, Giulia Farina, Fatih Aksu et al.

Major pathological response (pR) following neoadjuvant therapy is a clinically meaningful endpoint in non-small cell lung cancer, strongly associated with improved survival. However, accurate preoperative prediction of pR remains challenging, particularly in real-world clinical settings characterized by limited data availability and incomplete clinical profiles. In this study, we propose a multimodal deep learning framework designed to address these constraints by integrating foundation model-based CT feature extraction with a missing-aware architecture for clinical variables. This approach enables robust learning from small cohorts while explicitly modeling missing clinical information, without relying on conventional imputation strategies. A weighted fusion mechanism is employed to leverage the complementary contributions of imaging and clinical modalities, yielding a multimodal model that consistently outperforms both unimodal imaging and clinical baselines. These findings underscore the added value of integrating heterogeneous data sources and highlight the potential of multimodal, missing-aware systems to support pR prediction under realistic clinical conditions.

Massimiliano Mantegna, Elena Mulero Ayllón, Alice Natalina Caragliano et al.

Predicting tumor evolution during radiotherapy is a clinically critical challenge, particularly when longitudinal changes are driven by both anatomy and treatment. In this work, we introduce a Virtual Treatment (VT) framework that formulates non-small cell lung cancer (NSCLC) progression as a dose-aware multimodal conditional image-to-image translation problem. Given a CT scan, baseline clinical variables, and a specified radiation dose increment, VT aims to synthesize plausible follow-up CT images reflecting treatment-induced anatomical changes. We evaluate the proposed framework on a longitudinal dataset of 222 stage III NSCLC patients, comprising 895 CT scans acquired during radiotherapy under irregular clinical schedules. The generative process is conditioned on delivered dose increments together with demographic and tumor-related clinical variables. Representative GAN-based and diffusion-based models are benchmarked across 2D and 2.5D configurations. Quantitative and qualitative results indicate that diffusion-based models benefit more consistently from multimodal, dose-aware conditioning and produce more stable and anatomically plausible tumor evolution trajectories than GAN-based baselines, supporting the potential of VT as a tool for in-silico treatment monitoring and adaptive radiotherapy research in NSCLC.

Marco Salmè, Federico Siciliano, Fabrizio Silvestri et al.

Radiology Report Generation (RRG) through Vision-Language Models (VLMs) promises to reduce documentation burden, improve reporting consistency, and accelerate clinical workflows. However, their clinical adoption remains limited by the lack of interpretability and the tendency to hallucinate findings misaligned with imaging evidence. Existing research typically treats interpretability and accuracy as separate objectives, with concept-based explainability techniques focusing primarily on transparency, while Retrieval-Augmented Generation (RAG) methods targeting factual grounding through external retrieval. We present Concept-Enhanced Multimodal RAG (CEMRAG), a unified framework that decomposes visual representations into interpretable clinical concepts and integrates them with multimodal RAG. This approach exploits enriched contextual prompts for RRG, improving both interpretability and factual accuracy. Experiments on MIMIC-CXR and IU X-Ray across multiple VLM architectures, training regimes, and retrieval configurations demonstrate consistent improvements over both conventional RAG and concept-only baselines on clinical accuracy metrics and standard NLP measures. These results challenge the assumed trade-off between interpretability and performance, showing that transparent visual concepts can enhance rather than compromise diagnostic accuracy in medical VLMs. Our modular design decomposes interpretability into visual transparency and structured language model conditioning, providing a principled pathway toward clinically trustworthy AI-assisted radiology.

Filippo Ruffini, Camillo Maria Caruso, Claudia Tacconi et al.

Accurate survival prediction in Non-Small Cell Lung Cancer (NSCLC) requires the integration of heterogeneous clinical, radiological, and histopathological information. While Multimodal Deep Learning (MDL) offers a promises for precision prognosis and survival prediction, its clinical applicability is severely limited by small cohort sizes and the presence of missing modalities, often forcing complete-case filtering or aggressive imputation. In this work, we present a missing-aware multimodal survival framework that integrates Computed Tomography (CT), Whole-Slide Histopathology (WSI) Images, and structured clinical variables for overall survival modeling in unresectable stage II-III NSCLC. By leveraging Foundation Models (FM) for modality-specific feature extraction and a missing-aware encoding strategy, the proposed approach enables intermediate multimodal fusion under naturally incomplete modality profiles. The proposed architecture is resilient to missing modalities by design, allowing the model to utilize all available data without being forced to drop patients during training or inference. Experimental results demonstrate that intermediate fusion consistently outperforms unimodal baselines as well as early and late fusion strategies, with the strongest performance achieved by the fusion of WSI and clinical modalities (73.30 C-index). Further analyses of modality importance reveal an adaptive behavior in which less informative modalities, i.e., CT modality, are automatically down-weighted and contribute less to the final survival prediction.

Omar Coser, Loredana Zollo, Paolo Soda et al.

Amos et al. (2024) showed that the accuracy of Transformer models in sequence classification can be significantly improved by first pretraining with a masked token prediction objective without external data or augmentation, a procedure referred to as self-pretraining (SPT). While the primary objective of Amos et al. (2024) was to showcase that Transformers can achieve strong performance on the Long-Range Arena (LRA), their pipeline raises more fundamental questions: How does SPT drive optimization to better solutions? Why can standard supervised training fail in Transformers? To better understand this, we replicate and systematically ablate the findings of Amos et al. (2024). Our ablations suggest that a central bottleneck in the studied settings is not depth or generalization alone, but the ability of label supervision to learn useful query-key Attention patterns from random initialization. With a minimal setup, we identify learning proximity interactions - turning absolute positional encodings into proximity-biased Attention scores - as a key source of the improvements brought by SPT. Finally, in a simplified theoretical setup, we show that label supervision can be locally blind to certain Attention-score directions that are instead detectable through masked reconstruction.

Omar Coser, Christian Tamantini, Matteo Tortora et al.

Wearable robotics for lower-limb assistance have become a pivotal area of research, aiming to enhance mobility for individuals with physical impairments or augment the performance of able-bodied users. Accurate and adaptive control systems are essential to ensure seamless interaction between the wearer and the robotic device, particularly when navigating diverse and dynamic terrains. Despite the recent advances in neural networks for time series analysis, no attempts have been directed towards the classification of ground conditions, categorized into five classes and subsequently determining the ramp's slope and stair's height. In this respect, this paper presents an experimental comparison between eight deep neural network backbones to predict high-level locomotion parameters across diverse terrains. All the models are trained on the publicly available CAMARGO 2021 dataset. IMU-only data equally or outperformed IMU+EMG inputs, promoting a cost-effective and efficient design. Indeeds, using three IMU sensors, the LSTM achieved high terrain classification accuracy (0.94 +- 0.04) and precise ramp slope (1.95 +- 0.58°) and the CNN-LSTM a stair height (15.65 +- 7.40 mm) estimations. As a further contribution, SHAP analysis justified sensor reduction without performance loss, ensuring a lightweight setup. The system operates with ~2 ms inference time, supporting real-time applications. The code is code available at https://github.com/cosbidev/Human-Locomotion-Identification.

Francesco Di Feola, Lorenzo Tronchin, Valerio Guarrasi et al.

Generative Adversarial Networks (GANs) have proved as a powerful framework for denoising applications in medical imaging. However, GAN-based denoising algorithms still suffer from limitations in capturing complex relationships within the images. In this regard, the loss function plays a crucial role in guiding the image generation process, encompassing how much a synthetic image differs from a real image. To grasp highly complex and non-linear textural relationships in the training process, this work presents a novel approach to capture and embed multi-scale texture information into the loss function. Our method introduces a differentiable multi-scale texture representation of the images dynamically aggregated by a self-attention layer, thus exploiting end-to-end gradient-based optimization. We validate our approach by carrying out extensive experiments in the context of low-dose CT denoising, a challenging application that aims to enhance the quality of noisy CT scans. We utilize three publicly available datasets, including one simulated and two real datasets. The results are promising as compared to other well-established loss functions, being also consistent across three different GAN architectures. The code is available at: https://github.com/TrainLaboratory/MultiScaleTextureLoss-MSTLF

Daniele Molino, Camillo Maria Caruso, Filippo Ruffini et al.

Objective: While recent advances in text-conditioned generative models have enabled the synthesis of realistic medical images, progress has been largely confined to 2D modalities such as chest X-rays. Extending text-to-image generation to volumetric CT remains a significant challenge, due to its high dimensionality, anatomical complexity, and the absence of robust frameworks that align vision-language data in 3D medical imaging. Methods: We introduce a novel architecture for Text-to-CT generation that combines a latent diffusion model with a 3D contrastive vision-language pretraining scheme. Our approach leverages a dual-encoder CLIP-style model trained on paired CT volumes and radiology reports to establish a shared embedding space, which serves as the conditioning input for generation. CT volumes are compressed into a low-dimensional latent space via a pretrained volumetric VAE, enabling efficient 3D denoising diffusion without requiring external super-resolution stages. Results: We evaluate our method on the CT-RATE dataset and conduct a comprehensive assessment of image fidelity, clinical relevance, and semantic alignment. Our model achieves competitive performance across all tasks, significantly outperforming prior baselines for text-to-CT generation. Moreover, we demonstrate that CT scans synthesized by our framework can effectively augment real data, improving downstream diagnostic performance. Conclusion: Our results show that modality-specific vision-language alignment is a key component for high-quality 3D medical image generation. By integrating contrastive pretraining and volumetric diffusion, our method offers a scalable and controllable solution for synthesizing clinically meaningful CT volumes from text, paving the way for new applications in data augmentation, medical education, and automated clinical simulation. Code at https://github.com/cosbidev/Text2CT.

Giulia Romoli, Alessia Capoccia, Filippo Ruffini et al.

Medical image-to-image (I2I) translation enables virtual scanning, i.e. the synthesis of a target imaging modality from a source one without additional acquisitions. Despite growing interest, most proposed methods operate on 2D slices, are evaluated on isolated tasks with different experimental set-ups and lack clinical validation. The primary contribution of this work is a reproducible, standardized comparative evaluation of 3D I2I translation methods in oncological imaging, designed to standardize preprocessing, splitting, inference, and multi-level evaluation across heterogeneous clinical tasks. Within this framework, we compare seven generative models, three Generative Adversarial Networks (GANs: Pix2Pix, CycleGAN, SRGAN) and four latent generative models (Latent Diffusion Model, Latent Diffusion Model+ControlNet, Brownian Bridge, Flow Matching), across eleven datasets spanning three anatomical regions (head/neck, lung, pelvis) and four translation directions (cone-beam CT to CT, MRI to CT, CT to PET, MRI T2-weighted to T2-FLAIR), for a total of 77 experiments under uniform training, inference, and evaluation conditions. The results show that GANs outperform latent generative models across all tasks, with SRGAN achieving statistically significant superiority. Our lesion-level analysis reveals that all models struggle with small lesions and that, in CT to PET synthesis, models reproduce lesion shape more reliably than absolute uptake-related intensity. We also performed a Visual Turing test administered to 17 physicians, including 15 radiologists, which shows near-chance classification accuracy (56.7%), confirming that synthetic volumes are largely indistinguishable from real acquisitions, while exposing a dissociation between quantitative metrics and clinical preference.

Camillo Maria Caruso, Valerio Guarrasi, Paolo Soda

Multimodal deep learning has shown strong potential in medical applications by integrating heterogeneous data sources such as medical images and structured clinical variables. However, most existing approaches implicitly assume complete modality availability, an assumption that rarely holds in real-world clinical settings where entire modalities and individual features are frequently missing. In this work, we propose a multimodal transformer framework for joint vision-tabular learning explicitly designed to operate under pervasive modality missingness, without relying on imputation or heuristic model switching. The architecture integrates three components: a vision, a tabular, and a multimodal fusion encoder. Unimodal representations are weighted through learnable modality tokens and fused via intermediate fusion with masked self-attention, which excludes missing tokens and modalities from information aggregation and gradient propagation. To further enhance resilience, we introduce a modality-dropout regularization strategy that stochastically removes available modalities during training, encouraging the model to exploit complementary information under partial data availability. We evaluate our approach on the MIMIC-CXR dataset paired with structured clinical data from MIMIC-IV for multilabel classification of 14 diagnostic findings with incomplete annotations. Two parallel systematic stress-test protocols progressively increase training and inference missingness in each modality separately, spanning fully multimodal to fully unimodal scenarios. Across all missingness regimes, the proposed method consistently outperforms representative baselines, showing smoother performance degradation and improved robustness. Ablation studies further demonstrate that attention-level masking and intermediate fusion with joint fine-tuning are key to resilient multimodal inference.

Guido Manni, Clemente Lauretti, Loredana Zollo et al.

Deep learning has revolutionized medical imaging, but its effectiveness is severely limited by insufficient labeled training data. This paper introduces a novel GAN-based semi-supervised learning framework specifically designed for low labeled-data regimes, evaluated across settings with 5 to 50 labeled samples per class. Our approach integrates three specialized neural networks -- a generator for class-conditioned image translation, a discriminator for authenticity assessment and classification, and a dedicated classifier -- within a three-phase training framework. The method alternates between supervised training on limited labeled data and unsupervised learning that leverages abundant unlabeled images through image-to-image translation rather than generation from noise. We employ ensemble-based pseudo-labeling that combines confidence-weighted predictions from the discriminator and classifier with temporal consistency through exponential moving averaging, enabling reliable label estimation for unlabeled data. Comprehensive evaluation across eleven MedMNIST datasets demonstrates that our approach achieves statistically significant improvements over six state-of-the-art GAN-based semi-supervised methods, with particularly strong performance in the extreme 5-shot setting where the scarcity of labeled data is most challenging. The framework maintains its superiority across all evaluated settings (5, 10, 20, and 50 shots per class). Our approach offers a practical solution for medical imaging applications where annotation costs are prohibitive, enabling robust classification performance even with minimal labeled data. Code is available at https://github.com/GuidoManni/SPARSE.

Irene Iele, Francesco Di Feola, Valerio Guarrasi et al.

Image-to-image translation has emerged as a powerful technique in medical imaging, enabling tasks such as image denoising and cross-modality conversion. However, it suffers from limitations in handling out-of-distribution samples without causing performance degradation. To address this limitation, we propose a novel Test-Time Adaptation (TTA) framework that dynamically adjusts the translation process based on the characteristics of each test sample. Our method introduces a Reconstruction Module to quantify the domain shift and a Dynamic Adaptation Block that selectively modifies the internal features of a pretrained translation model to mitigate the shift without compromising the performance on in-distribution samples that do not require adaptation. We evaluate our approach on two medical image-to-image translation tasks: low-dose CT denoising and T1 to T2 MRI translation, showing consistent improvements over both the baseline translation model without TTA and prior TTA methods. Our analysis highlights the limitations of the state-of-the-art that uniformly apply the adaptation to both out-of-distribution and in-distribution samples, demonstrating that dynamic, sample-specific adjustment offers a promising path to improve model resilience in real-world scenarios. The code is available at: https://github.com/Sample-Aware-TTA/Code.

Alice Natalina Caragliano, Valerio Guarrasi, Michela Gravina et al.

Pathological complete response (pCR) is a key prognostic factor in breast cancer patients undergoing neoadjuvant therapy, strongly associated with long-term survival and treatment personalization. However, accurate pre-treatment pCR prediction remains challenging due to severe class imbalance and limited generalizability across diverse clinical settings. In this work, we propose a multimodal stepwise clinically-guided attention learning framework for pCR prediction from breast magnetic resonance imaging (MRI), designed to address these limitations through medically grounded spatial guidance and multimodal integration. The approach follows a stepwise training strategy inspired by physician reasoning: the model first learns global discriminative imaging patterns, then attention mechanisms are introduced to constrain the network toward tumor regions, and finally clinical variables are integrated to refine decision-making. This guidance strategy encourages prioritization of task-relevant features, improving identification of responders despite their limited representation in the dataset. Moreover, grounding attention in anatomically consistent tumor regions reduces reliance on dataset-specific patterns, thereby enhancing cross-institutional generalization. The framework is evaluated through external validation across heterogeneous MRI cohorts. Compared to non-guided single-stage baselines, the proposed approach improves sensitivity while maintaining competitive specificity, and produces anatomically coherent attention maps that support interpretation of the model's predictions. These findings highlight the potential of clinically-guided multimodal attention learning for robust and generalizable pCR prediction in breast cancer.

Fatih Aksu, Laura Ciuffetti, Francesco Di Feola et al.

Accurate histological differentiation between adenocarcinoma (ADC) and squamous cell carcinoma (SCC) is critical for personalized treatment in non-small cell lung cancer (NSCLC). While [$^{18}$F]FDG PET/CT is a standard tool for the clinical evaluation of lung cancer, its utility is often limited by high costs and radiation exposure. In this paper, we investigate the feasibility of "virtual scanning" as a feature-enhancement strategy by evaluating whether synthetic PET data can provide complementary feature representations to supplement anatomical CT scans in histological subtype classification. We propose a framework that leverages a 3D Pix2Pix Generative Adversarial Network (GAN), pretrained on the FDG-PET/CT Lesions dataset, to synthesize pseudo-PET volumes from anatomical CT scans. These synthetic volumes are integrated with structural CT data within the MINT framework, a multi-stage intermediate fusion architecture. Our experiments, conducted on a multi-center dataset of 714 subjects, demonstrate that the inclusion of synthetic metabolic features significantly improves classification performance over a CT-only baseline. The multimodal approach achieved a statistically significant increase in the Area Under the Curve (AUC) from 0.489 to 0.591 and improved the Geometric Mean (GMean) from 0.305 to 0.524. These results suggest that synthetic PET scans provide discriminatory metabolic cues that enable deep learning models to exploit complementary cross-modal information, offering a potential feature-enhancement strategy for clinical scenarios where physical PET scans are unavailable.

Filippo Ruffini, Lorenzo Tronchin, Zhuoru Wu et al.

In the fight against the COVID-19 pandemic, leveraging artificial intelligence to predict disease outcomes from chest radiographic images represents a significant scientific aim. The challenge, however, lies in the scarcity of large, labeled datasets with compatible tasks for training deep learning models without leading to overfitting. Addressing this issue, we introduce a novel multi-dataset multi-task training framework that predicts COVID-19 prognostic outcomes from chest X-rays (CXR) by integrating correlated datasets from disparate sources, distant from conventional multi-task learning approaches, which rely on datasets with multiple and correlated labeling schemes. Our framework hypothesizes that assessing severity scores enhances the model's ability to classify prognostic severity groups, thereby improving its robustness and predictive power. The proposed architecture comprises a deep convolutional network that receives inputs from two publicly available CXR datasets, AIforCOVID for severity prognostic prediction and BRIXIA for severity score assessment, and branches into task-specific fully connected output networks. Moreover, we propose a multi-task loss function, incorporating an indicator function, to exploit multi-dataset integration. The effectiveness and robustness of the proposed approach are demonstrated through significant performance improvements in prognosis classification tasks across 18 different convolutional neural network backbones in different evaluation strategies. This improvement is evident over single-task baselines and standard transfer learning strategies, supported by extensive statistical analysis, showing great application potential.

Arianna Francesconi, Lazzaro di Biase, Donato Cappetta et al.

Alzheimer's disease (AD) poses significant global health challenges due to its increasing prevalence and associated societal costs. Early detection and diagnosis of AD are critical for delaying progression and improving patient outcomes. Traditional diagnostic methods and single-modality data often fall short in identifying early-stage AD and distinguishing it from Mild Cognitive Impairment (MCI). This study addresses these challenges by introducing a novel approach: multImodal enseMble via class BALancing diversity for iMbalancEd Data (IMBALMED). IMBALMED integrates multimodal data from the Alzheimer's Disease Neuroimaging Initiative database, including clinical assessments, neuroimaging phenotypes, biospecimen and subject characteristics data. It employs an ensemble of model classifiers, each trained with different class balancing techniques, to overcome class imbalance and enhance model accuracy. We evaluate IMBALMED on two diagnostic tasks (binary and ternary classification) and four binary early detection tasks (at 12, 24, 36, and 48 months), comparing its performance with state-of-the-art algorithms and an unbalanced dataset method. IMBALMED demonstrates superior diagnostic accuracy and predictive performance in both binary and ternary classification tasks, significantly improving early detection of MCI at 48-month time point. The method shows improved classification performance and robustness, offering a promising solution for early detection and management of AD.

Alice Natalina Caragliano, Filippo Ruffini, Carlo Greco et al.

Non-small cell lung cancer (NSCLC) remains a major global health challenge, with high post-surgical recurrence rates underscoring the need for accurate pathological response predictions to guide personalized treatments. Although artificial intelligence models show promise in this domain, their clinical adoption is limited by the lack of medically grounded guidance during training, often resulting in non-explainable intrinsic predictions. To address this, we propose Doctor-in-the-Loop, a novel framework that integrates expert-driven domain knowledge with explainable artificial intelligence techniques, directing the model toward clinically relevant anatomical regions and improving both interpretability and trustworthiness. Our approach employs a gradual multi-view strategy, progressively refining the model's focus from broad contextual features to finer, lesion-specific details. By incorporating domain insights at every stage, we enhance predictive accuracy while ensuring that the model's decision-making process aligns more closely with clinical reasoning. Evaluated on a dataset of NSCLC patients, Doctor-in-the-Loop delivers promising predictive performance and provides transparent, justifiable outputs, representing a significant step toward clinically explainable artificial intelligence in oncology.

Camillo Maria Caruso, Paolo Soda, Valerio Guarrasi

In healthcare, the integration of multimodal data is pivotal for developing comprehensive diagnostic and predictive models. However, managing missing data remains a significant challenge in real-world applications. We introduce MARIA (Multimodal Attention Resilient to Incomplete datA), a novel transformer-based deep learning model designed to address these challenges through an intermediate fusion strategy. Unlike conventional approaches that depend on imputation, MARIA utilizes a masked self-attention mechanism, which processes only the available data without generating synthetic values. This approach enables it to effectively handle incomplete datasets, enhancing robustness and minimizing biases introduced by imputation methods. We evaluated MARIA against 10 state-of-the-art machine learning and deep learning models across 8 diagnostic and prognostic tasks. The results demonstrate that MARIA outperforms existing methods in terms of performance and resilience to varying levels of data incompleteness, underscoring its potential for critical healthcare applications.

Filippo Ruffini, Elena Mulero Ayllon, Linlin Shen et al.

Despite the significant potential of Foundation Models (FMs) in medical imaging, their application to prognosis prediction remains challenging due to data scarcity, class imbalance, and task complexity, which limit their clinical adoption. This study introduces the first structured benchmark to assess the robustness and efficiency of transfer learning strategies for FMs compared with convolutional neural networks (CNNs) in predicting COVID-19 patient outcomes from chest X-rays. The goal is to systematically compare finetuning strategies, both classical and parameter efficient, under realistic clinical constraints related to data scarcity and class imbalance, offering empirical guidance for AI deployment in clinical workflows. Four publicly available COVID-19 chest X-ray datasets were used, covering mortality, severity, and ICU admission, with varying sample sizes and class imbalances. CNNs pretrained on ImageNet and FMs pretrained on general or biomedical datasets were adapted using full finetuning, linear probing, and parameter-efficient methods. Models were evaluated under full data and few shot regimes using the Matthews Correlation Coefficient (MCC) and Precision Recall AUC (PR-AUC), with cross validation and class weighted losses. CNNs with full fine-tuning performed robustly on small, imbalanced datasets, while FMs with Parameter-Efficient Fine-Tuning (PEFT), particularly LoRA and BitFit, achieved competitive results on larger datasets. Severe class imbalance degraded PEFT performance, whereas balanced data mitigated this effect. In few-shot settings, FMs showed limited generalization, with linear probing yielding the most stable results. No single fine-tuning strategy proved universally optimal: CNNs remain dependable for low-resource scenarios, whereas FMs benefit from parameter-efficient methods when data are sufficient.

Daniele Molino, Francesco di Feola, Linlin Shen et al.

Generative models have revolutionized Artificial Intelligence (AI), particularly in multimodal applications. However, adapting these models to the medical domain poses unique challenges due to the complexity of medical data and the stringent need for clinical accuracy. In this work, we introduce a framework specifically designed for multimodal medical data generation. By enabling the generation of multi-view chest X-rays and their associated clinical report, it bridges the gap between general-purpose vision-language models and the specialized requirements of healthcare. Leveraging the MIMIC-CXR dataset, the proposed framework shows superior performance in generating high-fidelity images and semantically coherent reports. Our quantitative evaluation reveals significant results in terms of FID and BLEU scores, showcasing the quality of the generated data. Notably, our framework achieves comparable or even superior performance compared to real data on downstream disease classification tasks, underlining its potential as a tool for medical research and diagnostics. This study highlights the importance of domain-specific adaptations in enhancing the relevance and utility of generative models for clinical applications, paving the way for future advancements in synthetic multimodal medical data generation.

Aurora Rofena, Claudia Lucia Piccolo, Bruno Beomonte Zobel et al.

Full-Field Digital Mammography (FFDM) is the primary imaging modality for routine breast cancer screening; however, its effectiveness is limited in patients with dense breast tissue or fibrocystic conditions. Contrast-Enhanced Spectral Mammography (CESM), a second-level imaging technique, offers enhanced accuracy in tumor detection. Nonetheless, its application is restricted due to higher radiation exposure, the use of contrast agents, and limited accessibility. As a result, CESM is typically reserved for select cases, leaving many patients to rely solely on FFDM despite the superior diagnostic performance of CESM. While biopsy remains the gold standard for definitive diagnosis, it is an invasive procedure that can cause discomfort for patients. We introduce a multimodal, multi-view deep learning approach for virtual biopsy, integrating FFDM and CESM modalities in craniocaudal and mediolateral oblique views to classify lesions as malignant or benign. To address the challenge of missing CESM data, we leverage generative artificial intelligence to impute CESM images from FFDM scans. Experimental results demonstrate that incorporating the CESM modality is crucial to enhance the performance of virtual biopsy. When real CESM data is missing, synthetic CESM images proved effective, outperforming the use of FFDM alone, particularly in multimodal configurations that combine FFDM and CESM modalities. The proposed approach has the potential to improve diagnostic workflows, providing clinicians with augmented intelligence tools to improve diagnostic accuracy and patient care. Additionally, as a contribution to the research community, we publicly release the dataset used in our experiments, facilitating further advancements in this field.

Fatih Aksu, Fabrizia Gelardi, Arturo Chiti et al.

Accurate classification of histological subtypes of non-small cell lung cancer (NSCLC) is essential in the era of precision medicine, yet current invasive techniques are not always feasible and may lead to clinical complications. This study presents a multi-stage intermediate fusion approach to classify NSCLC subtypes from CT and PET images. Our method integrates the two modalities at different stages of feature extraction, using voxel-wise fusion to exploit complementary information across varying abstraction levels while preserving spatial correlations. We compare our method against unimodal approaches using only CT or PET images to demonstrate the benefits of modality fusion, and further benchmark it against early and late fusion techniques to highlight the advantages of intermediate fusion during feature extraction. Additionally, we compare our model with the only existing intermediate fusion method for histological subtype classification using PET/CT images. Our results demonstrate that the proposed method outperforms all alternatives across key metrics, with an accuracy and AUC equal to 0.724 and 0.681, respectively. This non-invasive approach has the potential to significantly improve diagnostic accuracy, facilitate more informed treatment decisions, and advance personalized care in lung cancer management.

Aurora Rofena, Arianna Manchia, Claudia Lucia Piccolo et al.

Contrast-Enhanced Spectral Mammography (CESM) is a dual-energy mammographic technique that improves lesion visibility through the administration of an iodinated contrast agent. It acquires both a low-energy image, comparable to standard mammography, and a high-energy image, which are then combined to produce a dual-energy subtracted image highlighting lesion contrast enhancement. While CESM offers superior diagnostic accuracy compared to standard mammography, its use entails higher radiation exposure and potential side effects associated with the contrast medium. To address these limitations, we propose Seg-CycleGAN, a generative deep learning framework for Virtual Contrast Enhancement in CESM. The model synthesizes high-fidelity dual-energy subtracted images from low-energy images, leveraging lesion segmentation maps to guide the generative process and improve lesion reconstruction. Building upon the standard CycleGAN architecture, Seg-CycleGAN introduces localized loss terms focused on lesion areas, enhancing the synthesis of diagnostically relevant regions. Experiments on the CESM@UCBM dataset demonstrate that Seg-CycleGAN outperforms the baseline in terms of PSNR and SSIM, while maintaining competitive MSE and VIF. Qualitative evaluations further confirm improved lesion fidelity in the generated images. These results suggest that segmentation-aware generative models offer a viable pathway toward contrast-free CESM alternatives.

Elena Mulero Ayllón, Massimiliano Mantegna, Linlin Shen et al.

Accurate lung tumor segmentation is crucial for improving diagnosis, treatment planning, and patient outcomes in oncology. However, the complexity of tumor morphology, size, and location poses significant challenges for automated segmentation. This study presents a comprehensive benchmarking analysis of deep learning-based segmentation models, comparing traditional architectures such as U-Net and DeepLabV3, self-configuring models like nnUNet, and foundation models like MedSAM, and MedSAM~2. Evaluating performance across two lung tumor segmentation datasets, we assess segmentation accuracy and computational efficiency under various learning paradigms, including few-shot learning and fine-tuning. The results reveal that while traditional models struggle with tumor delineation, foundation models, particularly MedSAM~2, outperform them in both accuracy and computational efficiency. These findings underscore the potential of foundation models for lung tumor segmentation, highlighting their applicability in improving clinical workflows and patient outcomes.

Marco Salmè, Lorenzo Tronchin, Rosa Sicilia et al.

Data scarcity remains a critical bottleneck impeding technological advancements across various domains, including but not limited to medicine and precision agriculture. To address this challenge, we explore the potential of Deep Generative Models (DGMs) in producing synthetic data that satisfies the Generative Learning Trilemma: fidelity, diversity, and sampling efficiency. However, recognizing that these criteria alone are insufficient for practical applications, we extend the trilemma to include utility, robustness, and privacy, factors crucial for ensuring the applicability of DGMs in real-world scenarios. Evaluating these metrics becomes particularly challenging in data-scarce environments, as DGMs traditionally rely on large datasets to perform optimally. This limitation is especially pronounced in domains like medicine and precision agriculture, where ensuring acceptable model performance under data constraints is vital. To address these challenges, we assess the Generative Learning Trilemma in data-scarcity settings using state-of-the-art evaluation metrics, comparing three prominent DGMs: Variational Autoencoders (VAEs), Generative Adversarial Networks (GANs), and Diffusion Models (DMs). Furthermore, we propose a comprehensive framework to assess utility, robustness, and privacy in synthetic data generated by DGMs. Our findings demonstrate varying strengths among DGMs, with each model exhibiting unique advantages based on the application context. This study broadens the scope of the Generative Learning Trilemma, aligning it with real-world demands and providing actionable guidance for selecting DGMs tailored to specific applications.

Lorenzo Tronchin, Tommy Löfstedt, Paolo Soda et al.

The advancement of generative AI, particularly in medical imaging, confronts the trilemma of ensuring high fidelity, diversity, and efficiency in synthetic data generation. While Generative Adversarial Networks (GANs) have shown promise across various applications, they still face challenges like mode collapse and insufficient coverage of real data distributions. This work explores the use of GAN ensembles to overcome these limitations, specifically in the context of medical imaging. By solving a multi-objective optimisation problem that balances fidelity and diversity, we propose a method for selecting an optimal ensemble of GANs tailored for medical data. The selected ensemble is capable of generating diverse synthetic medical images that are representative of true data distributions and computationally efficient. Each model in the ensemble brings a unique contribution, ensuring minimal redundancy. We conducted a comprehensive evaluation using three distinct medical datasets, testing 22 different GAN architectures with various loss functions and regularisation techniques. By sampling models at different training epochs, we crafted 110 unique configurations. The results highlight the capability of GAN ensembles to enhance the quality and utility of synthetic medical images, thereby improving the efficacy of downstream tasks such as diagnostic modelling.

Irene Iele, Floriano Caprio, Paolo Soda et al.

Forecasting physiological signals can support proactive monitoring and timely clinical intervention by anticipating critical changes in patient status. In this work, we address multivariate multi-horizon forecasting of physiological time series by jointly predicting heart rate, oxygen saturation, pulse rate, and respiratory rate at forecasting horizons of 15, 30, and 60 seconds. We propose a hybrid quantum-classical architecture that integrates a Variational Quantum Circuit (VQC) within a recurrent neural backbone. A GRU encoder summarizes the historical observation window into a latent representation, which is then projected into quantum angles used to parameterize the VQC. The quantum layer acts as a learnable non-linear feature mixer, modeling cross-variable interactions before the final prediction stage. We evaluate the proposed approach on the BIDMC PPG and Respiration dataset under a Leave-One-Patient-Out protocol. The results show competitive accuracy compared with classical and deep learning baselines, together with greater robustness to noise and missing inputs. These findings suggest that hybrid quantum layers can provide useful inductive biases for physiological time series forecasting in small-cohort clinical settings.

Arianna Francesconi, Donato Cappetta, Fabio Rebecchi et al.

Parkinson's disease (PD) presents a growing global challenge, affecting over 10 million individuals, with prevalence expected to double by 2040. Early diagnosis remains difficult due to the late emergence of motor symptoms and limitations of traditional clinical assessments. In this study, we propose a novel pipeline that leverages keystroke dynamics as a non-invasive and scalable biomarker for remote PD screening and telemonitoring. Our methodology involves three main stages: (i) preprocessing of data from four distinct datasets, extracting four temporal signals and addressing class imbalance through the comparison of three methods; (ii) pre-training eight state-of-the-art deep-learning architectures on the two largest datasets, optimizing temporal windowing, stride, and other hyperparameters; (iii) fine-tuning on an intermediate-sized dataset and performing external validation on a fourth, independent cohort. Our results demonstrate that hybrid convolutional-recurrent and transformer-based models achieve strong external validation performance, with AUC-ROC scores exceeding 90% and F1-Score over 70%. Notably, a temporal convolutional model attains an AUC-ROC of 91.14% in external validation, outperforming existing methods that rely solely on internal validation. These findings underscore the potential of keystroke dynamics as a reliable digital biomarker for PD, offering a promising avenue for early detection and continuous monitoring.

Valerio Guarrasi, Klara Mogensen, Sara Tassinari et al.

Multimodal deep learning harnesses diverse imaging modalities, such as MRI sequences, to enhance diagnostic accuracy in medical imaging. A key challenge is determining the optimal timing for integrating these modalities-specifically, identifying the network layers where fusion modules should be inserted. Current approaches often rely on manual tuning or exhaustive search, which are computationally expensive without any guarantee of converging to optimal results. We propose a sequential forward search algorithm that incrementally activates and evaluates candidate fusion modules at different layers of a multimodal network. At each step, the algorithm retrains from previously learned weights and compares validation loss to identify the best-performing configuration. This process systematically reduces the search space, enabling efficient identification of the optimal fusion timing without exhaustively testing all possible module placements. The approach is validated on two multimodal MRI datasets, each addressing different classification tasks. Our algorithm consistently identified configurations that outperformed unimodal baselines, late fusion, and a brute-force ensemble of all potential fusion placements. These architectures demonstrated superior accuracy, F-score, and specificity while maintaining competitive or improved AUC values. Furthermore, the sequential nature of the search significantly reduced computational overhead, making the optimization process more practical. By systematically determining the optimal timing to fuse imaging modalities, our method advances multimodal deep learning for medical imaging. It provides an efficient and robust framework for fusion optimization, paving the way for improved clinical decision-making and more adaptable, scalable architectures in medical AI applications.

Alice Natalina Caragliano, Claudia Tacconi, Carlo Greco et al.

This study proposes a novel approach combining Multimodal Deep Learning with intrinsic eXplainable Artificial Intelligence techniques to predict pathological response in non-small cell lung cancer patients undergoing neoadjuvant therapy. Due to the limitations of existing radiomics and unimodal deep learning approaches, we introduce an intermediate fusion strategy that integrates imaging and clinical data, enabling efficient interaction between data modalities. The proposed Multimodal Doctor-in-the-Loop method further enhances clinical relevance by embedding clinicians' domain knowledge directly into the training process, guiding the model's focus gradually from broader lung regions to specific lesions. Results demonstrate improved predictive accuracy and explainability, providing insights into optimal data integration strategies for clinical applications.

Marco Salmè, Rosa Sicilia, Paolo Soda et al.

The integration of artificial intelligence in healthcare has opened new horizons for improving medical diagnostics and patient care. However, challenges persist in developing systems capable of generating accurate and contextually relevant radiology reports, particularly in low-resource languages. In this study, we present a comprehensive benchmark to evaluate the performance of instruction-tuned Vision-Language Models (VLMs) in the specialized task of radiology report generation across three low-resource languages: Italian, German, and Spanish. Employing the LLaVA architectural framework, we conducted a systematic evaluation of pre-trained models utilizing general datasets, domain-specific datasets, and low-resource language-specific datasets. In light of the unavailability of models that possess prior knowledge of both the medical domain and low-resource languages, we analyzed various adaptations to determine the most effective approach for these contexts. The results revealed that language-specific models substantially outperformed both general and domain-specific models in generating radiology reports, emphasizing the critical role of linguistic adaptation. Additionally, models fine-tuned with medical terminology exhibited enhanced performance across all languages compared to models with generic knowledge, highlighting the importance of domain-specific training. We also explored the influence of the temperature parameter on the coherence of report generation, providing insights for optimal model settings. Our findings highlight the importance of tailored language and domain-specific training for improving the quality and accuracy of radiological reports in multilingual settings. This research not only advances our understanding of VLMs adaptability in healthcare but also points to significant avenues for future investigations into model tuning and language-specific adaptations.

Francesco Di Feola, Ludovica Pompilio, Cecilia Assolito et al.

Computed Tomography (CT) plays a pivotal role in medical diagnosis; however, variability across reconstruction kernels hinders data-driven approaches, such as deep learning models, from achieving reliable and generalized performance. To this end, CT data harmonization has emerged as a promising solution to minimize such non-biological variances by standardizing data across different sources or conditions. In this context, Generative Adversarial Networks (GANs) have proved to be a powerful framework for harmonization, framing it as a style-transfer problem. However, GAN-based approaches still face limitations in capturing complex relationships within the images, which are essential for effective harmonization. In this work, we propose a novel texture-aware StarGAN for CT data harmonization, enabling one-to-many translations across different reconstruction kernels. Although the StarGAN model has been successfully applied in other domains, its potential for CT data harmonization remains unexplored. Furthermore, our approach introduces a multi-scale texture loss function that embeds texture information across different spatial and angular scales into the harmonization process, effectively addressing kernel-induced texture variations. We conducted extensive experimentation on a publicly available dataset, utilizing a total of 48667 chest CT slices from 197 patients distributed over three different reconstruction kernels, demonstrating the superiority of our method over the baseline StarGAN.

Valerio Guarrasi, Francesco Di Feola, Rebecca Restivo et al.

Generating positron emission tomography (PET) images from computed tomography (CT) scans via deep learning offers a promising pathway to reduce radiation exposure and costs associated with PET imaging, improving patient care and accessibility to functional imaging. Whole-body image translation presents challenges due to anatomical heterogeneity, often limiting generalized models. We propose a framework that segments whole-body CT images into four regions-head, trunk, arms, and legs-and uses district-specific Generative Adversarial Networks (GANs) for tailored CT-to-PET translation. Synthetic PET images from each region are stitched together to reconstruct the whole-body scan. Comparisons with a baseline non-segmented GAN and experiments with Pix2Pix and CycleGAN architectures tested paired and unpaired scenarios. Quantitative evaluations at district, whole-body, and lesion levels demonstrated significant improvements with our district-specific GANs. Pix2Pix yielded superior metrics, ensuring precise, high-quality image synthesis. By addressing anatomical heterogeneity, this approach achieves state-of-the-art results in whole-body CT-to-PET translation. This methodology supports healthcare Digital Twins by enabling accurate virtual PET scans from CT data, creating virtual imaging representations to monitor, predict, and optimize health outcomes.

Daniele Molino, Francesco Di Feola, Eliodoro Faiella et al.

The adoption of Artificial Intelligence in medical imaging holds great promise, yet it remains hindered by challenges such as data scarcity, privacy concerns, and the need for robust multimodal integration. While recent advances in generative modeling have enabled high-quality synthetic data generation, existing approaches are often limited to unimodal, unidirectional synthesis and therefore lack the ability to jointly synthesize multiple modalities while preserving clinical consistency. To address this challenge, we introduce XGeM, a 6.77-billion-parameter multimodal generative model designed to support flexible, any-to-any synthesis between medical data modalities. XGeM constructs a shared latent space via contrastive learning and introduces a novel Multi-Prompt Training strategy, enabling conditioning on arbitrary subsets of input modalities. This design allows the model to adapt to heterogeneous clinical inputs and generate multiple outputs jointly, preserving both semantic and structural coherence. We extensively validate XGeM: first we benchmark it against five competitors on the MIMIC-CXR dataset, a state-of-the-art dataset for multi-view Chest X-ray and radiological report generation. Secondly, we perform a Visual Turing Test with expert radiologists to assess the realism and clinical relevance of the generated data, ensuring alignment with real-world scenarios. Finally, we show how XGeM can support key medical data challenges such as anonymization, class imbalance, and data scarcity, underscoring its utility as a foundation model for medical data synthesis. Project page is at https://cosbidev.github.io/XGeM/.

Chongsheng Zhang, Paolo Soda, Jingjun Bi et al.

In predictive tasks, real-world datasets often present different degrees of imbalanced (i.e., long-tailed or skewed) distributions. While the majority (the head) classes have sufficient samples, the minority (the tail) classes can be under-represented by a rather limited number of samples. Data pre-processing has been shown to be very effective in dealing with such problems. On one hand, data re-sampling is a common approach to tackling class imbalance. On the other hand, dimension reduction, which reduces the feature space, is a conventional technique for reducing noise and inconsistencies in a dataset. However, the possible synergy between feature selection and data re-sampling for high-performance imbalance classification has rarely been investigated before. To address this issue, we carry out a comprehensive empirical study on the joint influence of feature selection and re-sampling on two-class imbalance classification. Specifically, we study the performance of two opposite pipelines for imbalance classification by applying feature selection before or after data re-sampling. We conduct a large number of experiments, with a total of 9225 tests, on 52 publicly available datasets, using 9 feature selection methods, 6 re-sampling approaches for class imbalance learning, and 3 well-known classification algorithms. Experimental results show that there is no constant winner between the two pipelines; thus both of them should be considered to derive the best performing model for imbalance classification. We find that the performance of an imbalance classification model not only depends on the classifier adopted and the ratio between the number of majority and minority samples, but also depends on the ratio between the number of samples and features. Overall, this study should provide new reference value for researchers and practitioners in imbalance learning.

Paolo Soda, Natascha Claudia D'Amico, Jacopo Tessadori et al.

Recent epidemiological data report that worldwide more than 53 million people have been infected by SARS-CoV-2, resulting in 1.3 million deaths. The disease has been spreading very rapidly and few months after the identification of the first infected, shortage of hospital resources quickly became a problem. In this work we investigate whether chest X-ray (CXR) can be used as a possible tool for the early identification of patients at risk of severe outcome, like intensive care or death. CXR is a radiological technique that compared to computed tomography (CT) it is simpler, faster, more widespread and it induces lower radiation dose. We present a dataset including data collected from 820 patients by six Italian hospitals in spring 2020 during the first COVID-19 emergency. The dataset includes CXR images, several clinical attributes and clinical outcomes. We investigate the potential of artificial intelligence to predict the prognosis of such patients, distinguishing between severe and mild cases, thus offering a baseline reference for other researchers and practitioners. To this goal, we present three approaches that use features extracted from CXR images, either handcrafted or automatically by convolutional neuronal networks, which are then integrated with the clinical data. Exhaustive evaluation shows promising performance both in 10-fold and leave-one-centre-out cross-validation, implying that clinical data and images have the potential to provide useful information for the management of patients and hospital resources.