Pedro R. A. S. Bassi

Pedro R. A. S. Bassi, Haydr A. H. Ali, Andrea Cavalli et al.

Application of Artificial Intelligence (AI) in critical domains, like the medical one, is often hampered by shortcut learning, which hinders AI generalization to diverse hospitals and patients. Shortcut learning can be caused, for example, by background biases -- features in image backgrounds that are spuriously correlated to classification labels (e.g., words in X-rays). To mitigate the influence of image background and foreground bias on AI, we introduce a new training paradigm, dubbed Training with Explanations Alone (TEA). TEA trains a classifier (TEA student) only by making its explanation heatmaps match target heatmaps from a larger teacher model. By learning from its explanation heatmaps, the TEA student pays attention to the same image features as the teacher. For example, a teacher uses a large segmenter to remove image backgrounds before classification, thus ignoring background bias. By learning from the teacher's explanation heatmaps, the TEA student learns to also ignore backgrounds -- but it does not need a segmenter. With different teachers, the TEA student can also resist bias in the image foreground. Surprisingly, by training with heatmaps alone the student output naturally matches the teacher output -- with no loss function applied to the student output. We compared the TEA student against 14 state-of-the-art methods in 5 datasets with strong background or foreground bias, including Waterbirds and an X-Ray dataset for COVID-19 and pneumonia classification. The TEA student had better resistance to bias, strongly surpassing state-of-the-art methods, and generalizing better to hospitals not seen in training.

Pedro R. A. S. Bassi, Wenxuan Li, Yucheng Tang et al.

How can we test AI performance? This question seems trivial, but it isn't. Standard benchmarks often have problems such as in-distribution and small-size test sets, oversimplified metrics, unfair comparisons, and short-term outcome pressure. As a consequence, good performance on standard benchmarks does not guarantee success in real-world scenarios. To address these problems, we present Touchstone, a large-scale collaborative segmentation benchmark of 9 types of abdominal organs. This benchmark is based on 5,195 training CT scans from 76 hospitals around the world and 5,903 testing CT scans from 11 additional hospitals. This diverse test set enhances the statistical significance of benchmark results and rigorously evaluates AI algorithms across various out-of-distribution scenarios. We invited 14 inventors of 19 AI algorithms to train their algorithms, while our team, as a third party, independently evaluated these algorithms on three test sets. In addition, we also evaluated pre-existing AI frameworks--which, differing from algorithms, are more flexible and can support different algorithms--including MONAI from NVIDIA, nnU-Net from DKFZ, and numerous other open-source frameworks. We are committed to expanding this benchmark to encourage more innovation of AI algorithms for the medical domain.

Pedro R. A. S. Bassi, Wenxuan Li, Jieneng Chen et al.

Tumor segmentation in CT scans is key for diagnosis, surgery, and prognosis, yet segmentation masks are scarce because their creation requires time and expertise. Public abdominal CT datasets have from dozens to a couple thousand tumor masks, but hospitals have hundreds of thousands of tumor CTs with radiology reports. Thus, leveraging reports to improve segmentation is key for scaling. In this paper, we propose a report-supervision loss (R-Super) that converts radiology reports into voxel-wise supervision for tumor segmentation AI. We created a dataset with 6,718 CT-Report pairs (from the UCSF Hospital), and merged it with public CT-Mask datasets (from AbdomenAtlas 2.0). We used our R-Super to train with these masks and reports, and strongly improved tumor segmentation in internal and external validation--F1 Score increased by up to 16% with respect to training with masks only. By leveraging readily available radiology reports to supplement scarce segmentation masks, R-Super strongly improves AI performance both when very few training masks are available (e.g., 50), and when many masks were available (e.g., 1.7K). Project: https://github.com/MrGiovanni/R-Super

Yixiong Chen, Wenjie Xiao, Pedro R. A. S. Bassi et al.

Vision-Language Models (VLMs) have shown promise in various 2D visual tasks, yet their readiness for 3D clinical diagnosis remains unclear due to stringent demands for recognition precision, reasoning ability, and domain knowledge. To systematically evaluate these dimensions, we present DeepTumorVQA, a diagnostic visual question answering (VQA) benchmark targeting abdominal tumors in CT scans. It comprises 9,262 CT volumes (3.7M slices) from 17 public datasets, with 395K expert-level questions spanning four categories: Recognition, Measurement, Visual Reasoning, and Medical Reasoning. DeepTumorVQA introduces unique challenges, including small tumor detection and clinical reasoning across 3D anatomy. Benchmarking four advanced VLMs (RadFM, M3D, Merlin, CT-CHAT), we find current models perform adequately on measurement tasks but struggle with lesion recognition and reasoning, and are still not meeting clinical needs. Two key insights emerge: (1) large-scale multimodal pretraining plays a crucial role in DeepTumorVQA testing performance, making RadFM stand out among all VLMs. (2) Our dataset exposes critical differences in VLM components, where proper image preprocessing and design of vision modules significantly affect 3D perception. To facilitate medical multimodal research, we have released DeepTumorVQA as a rigorous benchmark: https://github.com/Schuture/DeepTumorVQA.

Pedro R. A. S. Bassi, Mehmet Can Yavuz, Kang Wang et al.

Cancers identified in CT scans are usually accompanied by detailed radiology reports, but publicly available CT datasets often lack these essential reports. This absence limits their usefulness for developing accurate report generation AI. To address this gap, we present AbdomenAtlas 3.0, the first public, high-quality abdominal CT dataset with detailed, expert-reviewed radiology reports. All reports are paired with per-voxel masks and they describe liver, kidney and pancreatic tumors. AbdomenAtlas 3.0 has 9,262 triplets of CT, mask and report--3,955 with tumors. These CT scans come from 17 public datasets. Besides creating the reports for these datasets, we expanded their number of tumor masks by 4.2x, identifying 3,011 new tumor cases. Notably, the reports in AbdomenAtlas 3.0 are more standardized, and generated faster than traditional human-made reports. They provide details like tumor size, location, attenuation and surgical resectability. These reports were created by 12 board-certified radiologists using our proposed RadGPT, a novel framework that converted radiologist-revised tumor segmentation masks into structured and narrative reports. Besides being a dataset creation tool, RadGPT can also become a fully-automatic, segmentation-assisted report generation method. We benchmarked this method and 5 state-of-the-art report generation vision-language models. Our results show that segmentation strongly improves tumor detection in AI-made reports.

Xinran Li, Yi Shuai, Chen Liu et al.

Tumor synthesis can generate examples that AI often misses or over-detects, improving AI performance by training on these challenging cases. However, existing synthesis methods, which are typically unconditional -- generating images from random variables -- or conditioned only by tumor shapes, lack controllability over specific tumor characteristics such as texture, heterogeneity, boundaries, and pathology type. As a result, the generated tumors may be overly similar or duplicates of existing training data, failing to effectively address AI's weaknesses. We propose a new text-driven tumor synthesis approach, termed TextoMorph, that provides textual control over tumor characteristics. This is particularly beneficial for examples that confuse the AI the most, such as early tumor detection (increasing Sensitivity by +8.5%), tumor segmentation for precise radiotherapy (increasing DSC by +6.3%), and classification between benign and malignant tumors (improving Sensitivity by +8.2%). By incorporating text mined from radiology reports into the synthesis process, we increase the variability and controllability of the synthetic tumors to target AI's failure cases more precisely. Moreover, TextoMorph uses contrastive learning across different texts and CT scans, significantly reducing dependence on scarce image-report pairs (only 141 pairs used in this study) by leveraging a large corpus of 34,035 radiology reports. Finally, we have developed rigorous tests to evaluate synthetic tumors, including Text-Driven Visual Turing Test and Radiomics Pattern Analysis, showing that our synthetic tumors is realistic and diverse in texture, heterogeneity, boundaries, and pathology.

Wenxuan Li, Pedro R. A. S. Bassi, Tianyu Lin et al.

Building trusted datasets is critical for transparent and responsible Medical AI (MAI) research, but creating even small, high-quality datasets can take years of effort from multidisciplinary teams. This process often delays AI benefits, as human-centric data creation and AI-centric model development are treated as separate, sequential steps. To overcome this, we propose ScaleMAI, an agent of AI-integrated data curation and annotation, allowing data quality and AI performance to improve in a self-reinforcing cycle and reducing development time from years to months. We adopt pancreatic tumor detection as an example. First, ScaleMAI progressively creates a dataset of 25,362 CT scans, including per-voxel annotations for benign/malignant tumors and 24 anatomical structures. Second, through progressive human-in-the-loop iterations, ScaleMAI provides Flagship AI Model that can approach the proficiency of expert annotators (30-year experience) in detecting pancreatic tumors. Flagship Model significantly outperforms models developed from smaller, fixed-quality datasets, with substantial gains in tumor detection (+14%), segmentation (+5%), and classification (72%) on three prestigious benchmarks. In summary, ScaleMAI transforms the speed, scale, and reliability of medical dataset creation, paving the way for a variety of impactful, data-driven applications.

Wenxuan Li, Xinze Zhou, Qi Chen et al.

PanTS is a large-scale, multi-institutional dataset curated to advance research in pancreatic CT analysis. It contains 36,390 CT scans from 145 medical centers, with expert-validated, voxel-wise annotations of over 993,000 anatomical structures, covering pancreatic tumors, pancreas head, body, and tail, and 24 surrounding anatomical structures such as vascular/skeletal structures and abdominal/thoracic organs. Each scan includes metadata such as patient age, sex, diagnosis, contrast phase, in-plane spacing, slice thickness, etc. AI models trained on PanTS achieve significantly better performance in pancreatic tumor detection, localization, and segmentation compared to those trained on existing public datasets. Our analysis indicates that these gains are directly attributable to the 16x larger-scale tumor annotations and indirectly supported by the 24 additional surrounding anatomical structures. As the largest and most comprehensive resource of its kind, PanTS offers a new benchmark for developing and evaluating AI models in pancreatic CT analysis.

Pedro R. A. S. Bassi, Qilong Wu, Wenxuan Li et al.

As medical datasets rapidly expand, creating detailed annotations of different body structures becomes increasingly expensive and time-consuming. We consider that requesting radiologists to create detailed annotations is unnecessarily burdensome and that pre-existing AI models can largely automate this process. Following the spirit don't use a sledgehammer on a nut, we find that, rather than creating annotations from scratch, radiologists only have to review and edit errors if the Best-AI Labels have mistakes. To obtain the Best-AI Labels among multiple AI Labels, we developed an automatic tool, called Label Critic, that can assess label quality through tireless pairwise comparisons. Extensive experiments demonstrate that, when incorporated with our developed Image-Prompt pairs, pre-existing Large Vision-Language Models (LVLM), trained on natural images and texts, achieve 96.5% accuracy when choosing the best label in a pair-wise comparison, without extra fine-tuning. By transforming the manual annotation task (30-60 min/scan) into an automatic comparison task (15 sec/scan), we effectively reduce the manual efforts required from radiologists by an order of magnitude. When the Best-AI Labels are sufficiently accurate (81% depending on body structures), they will be directly adopted as the gold-standard annotations for the dataset, with lower-quality AI Labels automatically discarded. Label Critic can also check the label quality of a single AI Label with 71.8% accuracy when no alternatives are available for comparison, prompting radiologists to review and edit if the estimated quality is low (19% depending on body structures).

Pedro R. A. S. Bassi, Sergio Decherchi, Andrea Cavalli

Bias or spurious correlations in image backgrounds can impact neural networks, causing shortcut learning (Clever Hans Effect) and hampering generalization to real-world data. ISNet, a recently introduced architecture, proposed the optimization of Layer-Wise Relevance Propagation (LRP, an explanation technique) heatmaps, to mitigate the influence of backgrounds on deep classifiers. However, ISNet's training time scales linearly with the number of classes in an application. Here, we propose reformulated architectures whose training time becomes independent from this number. Additionally, we introduce a concise and model-agnostic LRP implementation. We challenge the proposed architectures using synthetic background bias, and COVID-19 detection in chest X-rays, an application that commonly presents background bias. The networks hindered background attention and shortcut learning, surpassing multiple state-of-the-art models on out-of-distribution test datasets. Representing a potentially massive training speed improvement over ISNet, the proposed architectures introduce LRP optimization into a gamut of applications that the original model cannot feasibly handle.

Pedro R. A. S. Bassi, Sergio S. J. Dertkigil, Andrea Cavalli

Features in images' backgrounds can spuriously correlate with the images' classes, representing background bias. They can influence the classifier's decisions, causing shortcut learning (Clever Hans effect). The phenomenon generates deep neural networks (DNNs) that perform well on standard evaluation datasets but generalize poorly to real-world data. Layer-wise Relevance Propagation (LRP) explains DNNs' decisions. Here, we show that the optimization of LRP heatmaps can minimize the background bias influence on deep classifiers, hindering shortcut learning. By not increasing run-time computational cost, the approach is light and fast. Furthermore, it applies to virtually any classification architecture. After injecting synthetic bias in images' backgrounds, we compared our approach (dubbed ISNet) to eight state-of-the-art DNNs, quantitatively demonstrating its superior robustness to background bias. Mixed datasets are common for COVID-19 and tuberculosis classification with chest X-rays, fostering background bias. By focusing on the lungs, the ISNet reduced shortcut learning. Thus, its generalization performance on external (out-of-distribution) test databases significantly surpassed all implemented benchmark models.

Pedro R. A. S. Bassi, Romis Attux

Objective: To propose novel SSVEP classification methodologies using deep neural networks (DNNs) and improve performances in single-channel and user-independent brain-computer interfaces (BCIs) with small data lengths. Approach: We propose the utilization of filter banks (creating sub-band components of the EEG signal) in conjunction with DNNs. In this context, we created three different models: a recurrent neural network (FBRNN) analyzing the time domain, a 2D convolutional neural network (FBCNN-2D) processing complex spectrum features and a 3D convolutional neural network (FBCNN-3D) analyzing complex spectrograms, which we introduce in this study as possible input for SSVEP classification. We tested our neural networks on three open datasets and conceived them so as not to require calibration from the final user, simulating a user-independent BCI. Results: The DNNs with the filter banks surpassed the accuracy of similar networks without this preprocessing step by considerable margins, and they outperformed common SSVEP classification methods (SVM and FBCCA) by even higher margins. Conclusion and significance: Filter banks allow different types of deep neural networks to more efficiently analyze the harmonic components of SSVEP. Complex spectrograms carry more information than complex spectrum features and the magnitude spectrum, allowing the FBCNN-3D to surpass the other CNNs. The performances obtained in the challenging classification problems indicates a strong potential for the construction of portable, economical, fast and low-latency BCIs.

Pedro R. A. S. Bassi, Romis Attux

Purpose: we evaluated the generalization capability of deep neural networks (DNNs), trained to classify chest X-rays as Covid-19, normal or pneumonia, using a relatively small and mixed dataset. Methods: we proposed a DNN to perform lung segmentation and classification, stacking a segmentation module (U-Net), an original intermediate module and a classification module (DenseNet201). To evaluate generalization, we tested the DNN with an external dataset (from distinct localities) and used Bayesian inference to estimate probability distributions of performance metrics. Results: our DNN achieved 0.917 AUC on the external test dataset, and a DenseNet without segmentation, 0.906. Bayesian inference indicated mean accuracy of 76.1% and [0.695, 0.826] 95% HDI (highest density interval, which concentrates 95% of the metric's probability mass) with segmentation and, without segmentation, 71.7% and [0.646, 0.786]. Conclusion: employing a novel DNN evaluation technique, which uses LRP and Brixia scores, we discovered that areas where radiologists found strong Covid-19 symptoms are the most important for the stacked DNN classification. External validation showed smaller accuracies than internal, indicating difficulty in generalization, which is positively affected by segmentation. Finally, the performance in the external dataset and the analysis with LRP suggest that DNNs can be trained in small and mixed datasets and still successfully detect Covid-19.

Pedro R. A. S. Bassi, Willian Rampazzo, Romis Attux

Objective: We used deep convolutional neural networks (DCNNs) to classify electroencephalography (EEG) signals in a steady-state visually evoked potentials (SSVEP) based single-channel brain-computer interface (BCI), which does not require calibration on the user. Methods: EEG signals were converted to spectrograms and served as input to train DCNNs using the transfer learning technique. We also modified and applied a data augmentation method, SpecAugment, generally employed for speech recognition. Furthermore, for comparison purposes, we classified the SSVEP dataset using Support-vector machines (SVMs) and Filter Bank canonical correlation analysis (FBCCA). Results: Excluding the evaluated user's data from the fine-tuning process, we reached 82.2% mean test accuracy and 0.825 mean F1-Score on 35 subjects from an open dataset, using a small data length (0.5 s), only one electrode (Oz) and the DCNN with transfer learning, window slicing (WS) and SpecAugment's time masks. Conclusion: The DCNN results surpassed SVM and FBCCA performances, using a single electrode and a small data length. Transfer learning provided minimal accuracy change, but made training faster. SpecAugment created a small performance improvement and was successfully combined with WS, yielding higher accuracies. Significance: We present a new methodology to solve the problem of SSVEP classification using DCNNs. We also modified a speech recognition data augmentation technique and applied it to the context of BCIs. The presented methodology surpassed performances obtained with FBCCA and SVMs (more traditional SSVEP classification methods) in BCIs with small data lengths and one electrode. This type of BCI can be used to develop small and fast systems.