Sandy Engelhardt

Firas Khader, Gustav Mueller-Franzes, Soroosh Tayebi Arasteh et al.

Recent advances in computer vision have shown promising results in image generation. Diffusion probabilistic models in particular have generated realistic images from textual input, as demonstrated by DALL-E 2, Imagen and Stable Diffusion. However, their use in medicine, where image data typically comprises three-dimensional volumes, has not been systematically evaluated. Synthetic images may play a crucial role in privacy preserving artificial intelligence and can also be used to augment small datasets. Here we show that diffusion probabilistic models can synthesize high quality medical imaging data, which we show for Magnetic Resonance Images (MRI) and Computed Tomography (CT) images. We provide quantitative measurements of their performance through a reader study with two medical experts who rated the quality of the synthesized images in three categories: Realistic image appearance, anatomical correctness and consistency between slices. Furthermore, we demonstrate that synthetic images can be used in a self-supervised pre-training and improve the performance of breast segmentation models when data is scarce (dice score 0.91 vs. 0.95 without vs. with synthetic data). The code is publicly available on GitHub: https://github.com/FirasGit/medicaldiffusion.

Matthias Eisenmann, Annika Reinke, Vivienn Weru et al. · utoronto

The number of international benchmarking competitions is steadily increasing in various fields of machine learning (ML) research and practice. So far, however, little is known about the common practice as well as bottlenecks faced by the community in tackling the research questions posed. To shed light on the status quo of algorithm development in the specific field of biomedical imaging analysis, we designed an international survey that was issued to all participants of challenges conducted in conjunction with the IEEE ISBI 2021 and MICCAI 2021 conferences (80 competitions in total). The survey covered participants' expertise and working environments, their chosen strategies, as well as algorithm characteristics. A median of 72% challenge participants took part in the survey. According to our results, knowledge exchange was the primary incentive (70%) for participation, while the reception of prize money played only a minor role (16%). While a median of 80 working hours was spent on method development, a large portion of participants stated that they did not have enough time for method development (32%). 25% perceived the infrastructure to be a bottleneck. Overall, 94% of all solutions were deep learning-based. Of these, 84% were based on standard architectures. 43% of the respondents reported that the data samples (e.g., images) were too large to be processed at once. This was most commonly addressed by patch-based training (69%), downsampling (37%), and solving 3D analysis tasks as a series of 2D tasks. K-fold cross-validation on the training set was performed by only 37% of the participants and only 50% of the participants performed ensembling based on multiple identical models (61%) or heterogeneous models (39%). 48% of the respondents applied postprocessing steps.

Malte Tölle, Fernando Navarro, Sebastian Eble et al.

Federated learning is one popular paradigm to train a joint model in a distributed, privacy-preserving environment. But partial annotations pose an obstacle meaning that categories of labels are heterogeneous over clients. We propose to learn a joint backbone in a federated manner, while each site receives its own multi-label segmentation head. By using Bayesian techniques we observe that the different segmentation heads although only trained on the individual client's labels also learn information about the other labels not present at the respective site. This information is encoded in their predictive uncertainty. To obtain a final prediction we leverage this uncertainty and perform a weighted averaging of the ensemble of distributed segmentation heads, which allows us to segment "locally unknown" structures. With our method, which we refer to as FUNAvg, we are even on-par with the models trained and tested on the same dataset on average. The code is publicly available at https://github.com/Cardio-AI/FUNAvg.

Malte Tölle, Ullrich Köthe, Florian André et al.

Differential privacy (DP) has arisen as the gold standard in protecting an individual's privacy in datasets by adding calibrated noise to each data sample. While the application to categorical data is straightforward, its usability in the context of images has been limited. Contrary to categorical data the meaning of an image is inherent in the spatial correlation of neighboring pixels making the simple application of noise infeasible. Invertible Neural Networks (INN) have shown excellent generative performance while still providing the ability to quantify the exact likelihood. Their principle is based on transforming a complicated distribution into a simple one e.g. an image into a spherical Gaussian. We hypothesize that adding noise to the latent space of an INN can enable differentially private image modification. Manipulation of the latent space leads to a modified image while preserving important details. Further, by conditioning the INN on meta-data provided with the dataset we aim at leaving dimensions important for downstream tasks like classification untouched while altering other parts that potentially contain identifying information. We term our method content-aware differential privacy (CADP). We conduct experiments on publicly available benchmarking datasets as well as dedicated medical ones. In addition, we show the generalizability of our method to categorical data. The source code is publicly available at https://github.com/Cardio-AI/CADP.

Sven Koehler, Tarique Hussain, Hamza Hussain et al.

Cardiac magnetic resonance (CMR) sequences visualise the cardiac function voxel-wise over time. Simultaneously, deep learning-based deformable image registration is able to estimate discrete vector fields which warp one time step of a CMR sequence to the following in a self-supervised manner. However, despite the rich source of information included in these 3D+t vector fields, a standardised interpretation is challenging and the clinical applications remain limited so far. In this work, we show how to efficiently use a deformable vector field to describe the underlying dynamic process of a cardiac cycle in form of a derived 1D motion descriptor. Additionally, based on the expected cardiovascular physiological properties of a contracting or relaxing ventricle, we define a set of rules that enables the identification of five cardiovascular phases including the end-systole (ES) and end-diastole (ED) without the usage of labels. We evaluate the plausibility of the motion descriptor on two challenging multi-disease, -center, -scanner short-axis CMR datasets. First, by reporting quantitative measures such as the periodic frame difference for the extracted phases. Second, by comparing qualitatively the general pattern when we temporally resample and align the motion descriptors of all instances across both datasets. The average periodic frame difference for the ED, ES key phases of our approach is $0.80\pm{0.85}$, $0.69\pm{0.79}$ which is slightly better than the inter-observer variability ($1.07\pm{0.86}$, $0.91\pm{1.6}$) and the supervised baseline method ($1.18\pm{1.91}$, $1.21\pm{1.78}$). Code and labels will be made available on our GitHub repository. https://github.com/Cardio-AI/cmr-phase-detection

Lalith Sharan, Halvar Kelm, Gabriele Romano et al.

Multi-point tracking is a challenging task that involves detecting points in the scene and tracking them across a sequence of frames. Computing detection-based measures like the F-measure on a frame-by-frame basis is not sufficient to assess the overall performance, as it does not interpret performance in the temporal domain. The main evaluation metric available comes from Multi-object tracking (MOT) methods to benchmark performance on datasets such as KITTI with the recently proposed higher order tracking accuracy (HOTA) metric, which is capable of providing a better description of the performance over metrics such as MOTA, DetA, and IDF1. While the HOTA metric takes into account temporal associations, it does not provide a tailored means to analyse the spatial associations of a dataset in a multi-camera setup. Moreover, there are differences in evaluating the detection task for points when compared to objects (point distances vs. bounding box overlap). Therefore in this work, we propose a multi-view higher order tracking metric (mvHOTA) to determine the accuracy of multi-point (multi-instance and multi-class) tracking methods, while taking into account temporal and spatial associations.mvHOTA can be interpreted as the geometric mean of detection, temporal, and spatial associations, thereby providing equal weighting to each of the factors. We demonstrate the use of this metric to evaluate the tracking performance on an endoscopic point detection dataset from a previously organised surgical data science challenge. Furthermore, we compare with other adjusted MOT metrics for this use-case, discuss the properties of mvHOTA, and show how the proposed multi-view Association and the Occlusion index (OI) facilitate analysis of methods with respect to handling of occlusions. The code is available at https://github.com/Cardio-AI/mvhota.

Matthias Eisenmann, Annika Reinke, Vivienn Weru et al.

International benchmarking competitions have become fundamental for the comparative performance assessment of image analysis methods. However, little attention has been given to investigating what can be learnt from these competitions. Do they really generate scientific progress? What are common and successful participation strategies? What makes a solution superior to a competing method? To address this gap in the literature, we performed a multi-center study with all 80 competitions that were conducted in the scope of IEEE ISBI 2021 and MICCAI 2021. Statistical analyses performed based on comprehensive descriptions of the submitted algorithms linked to their rank as well as the underlying participation strategies revealed common characteristics of winning solutions. These typically include the use of multi-task learning (63%) and/or multi-stage pipelines (61%), and a focus on augmentation (100%), image preprocessing (97%), data curation (79%), and postprocessing (66%). The "typical" lead of a winning team is a computer scientist with a doctoral degree, five years of experience in biomedical image analysis, and four years of experience in deep learning. Two core general development strategies stood out for highly-ranked teams: the reflection of the metrics in the method design and the focus on analyzing and handling failure cases. According to the organizers, 43% of the winning algorithms exceeded the state of the art but only 11% completely solved the respective domain problem. The insights of our study could help researchers (1) improve algorithm development strategies when approaching new problems, and (2) focus on open research questions revealed by this work.

Malte Tölle, Philipp Garthe, Clemens Scherer et al.

Federated learning is a renowned technique for utilizing decentralized data while preserving privacy. However, real-world applications often face challenges like partially labeled datasets, where only a few locations have certain expert annotations, leaving large portions of unlabeled data unused. Leveraging these could enhance transformer architectures ability in regimes with small and diversely annotated sets. We conduct the largest federated cardiac CT analysis to date (n=8,104) in a real-world setting across eight hospitals. Our two-step semi-supervised strategy distills knowledge from task-specific CNNs into a transformer. First, CNNs predict on unlabeled data per label type and then the transformer learns from these predictions with label-specific heads. This improves predictive accuracy and enables simultaneous learning of all partial labels across the federation, and outperforms UNet-based models in generalizability on downstream tasks. Code and model weights are made openly available for leveraging future cardiac CT analysis.

Malte Tölle, Lukas Burger, Halvar Kelm et al.

Purpose: Federated training is often hindered by heterogeneous datasets due to divergent data storage options, inconsistent naming schemes, varied annotation procedures, and disparities in label quality. This is particularly evident in the emerging multi-modal learning paradigms, where dataset harmonization including a uniform data representation and filtering options are of paramount importance. Methods: DICOM structured reports enable the standardized linkage of arbitrary information beyond the imaging domain and can be used within Python deep learning pipelines with highdicom. Building on this, we developed an open platform for data integration and interactive filtering capabilities that simplifies the process of assembling multi-modal datasets. Results: In this study, we extend our prior work by showing its applicability to more and divergent data types, as well as streamlining datasets for federated training within an established consortium of eight university hospitals in Germany. We prove its concurrent filtering ability by creating harmonized multi-modal datasets across all locations for predicting the outcome after minimally invasive heart valve replacement. The data includes DICOM data (i.e. computed tomography images, electrocardiography scans) as well as annotations (i.e. calcification segmentations, pointsets and pacemaker dependency), and metadata (i.e. prosthesis and diagnoses). Conclusion: Structured reports bridge the traditional gap between imaging systems and information systems. Utilizing the inherent DICOM reference system arbitrary data types can be queried concurrently to create meaningful cohorts for clinical studies. The graphical interface as well as example structured report templates will be made publicly available.

Salman Ul Hassan Dar, Arman Ghanaat, Jannik Kahmann et al.

Generative latent diffusion models have been established as state-of-the-art in data generation. One promising application is generation of realistic synthetic medical imaging data for open data sharing without compromising patient privacy. Despite the promise, the capacity of such models to memorize sensitive patient training data and synthesize samples showing high resemblance to training data samples is relatively unexplored. Here, we assess the memorization capacity of 3D latent diffusion models on photon-counting coronary computed tomography angiography and knee magnetic resonance imaging datasets. To detect potential memorization of training samples, we utilize self-supervised models based on contrastive learning. Our results suggest that such latent diffusion models indeed memorize training data, and there is a dire need for devising strategies to mitigate memorization.

Marvin Seyfarth, Sarah Kaye Müller, Arman Ghanaat et al.

Latent diffusion models (LDMs) have recently achieved strong performance in 3D medical image synthesis. However, modalities like cine cardiac MRI (CMR), representing a temporally synchronized 3D volume across the cardiac cycle, add an additional dimension that most generative approaches do not model directly. Instead, they factorize space and time or enforce temporal consistency through auxiliary mechanisms such as anatomical masks. Such strategies introduce structural biases that may limit global context integration and lead to subtle spatiotemporal discontinuities or physiologically inconsistent cardiac dynamics. We investigate whether a unified 4D generative model can learn continuous cardiac dynamics without architectural factorization. We propose CardioDiT, a fully 4D latent diffusion framework for short-axis cine CMR synthesis based on diffusion transformers. A spatiotemporal VQ-VAE encodes 2D+t slices into compact latents, which a diffusion transformer then models jointly as complete 3D+t volumes, coupling space and time throughout the generative process. We evaluate CardioDiT on public CMR datasets and a larger private cohort, comparing it to baselines with progressively stronger spatiotemporal coupling. Results show improved inter-slice consistency, temporally coherent motion, and realistic cardiac function distributions, suggesting that explicit 4D modeling with a diffusion transformer provides a principled foundation for spatiotemporal cardiac image synthesis. Code and models trained on public data are available at https://github.com/Cardio-AI/cardiodit.

Georgii Kostiuchik, Lalith Sharan, Benedikt Mayer et al.

Purpose: Machine learning models can only be reliably evaluated if training, validation, and test data splits are representative and not affected by the absence of classes of interest. Surgical workflow and instrument recognition tasks are complicated in this manner, because of heavy data imbalances resulting from different lengths of phases and their erratic occurrences. Furthermore, the issue becomes difficult as sub-properties that help define phases, like instrument (co-)occurrence, are usually not considered when defining the split. We argue that such sub-properties must be equally considered. Methods: This work presents a publicly available data visualization tool that enables interactive exploration of dataset splits for surgical phase and instrument recognition. It focuses on the visualization of the occurrence of phases, phase transitions, instruments, and instrument combinations across sets. Particularly, it facilitates the assessment and identification of sub-optimal dataset splits. Results: We performed an analysis of common Cholec80 dataset splits using the proposed application and were able to uncover phase transitions and combinations of instruments that were not represented in one of the sets. Additionally, we outlined possible improvements to the splits. A user study with ten participants demonstrated the ability of participants to solve a selection of data exploration tasks using the proposed application. Conclusion: In highly unbalanced class distributions, special care should be taken with respect to the selection of an appropriate dataset split. Our interactive data visualization tool presents a promising approach for the assessment of dataset splits for surgical phase and instrument recognition. Evaluation results show that it can enhance the development of machine learning models. The application is available at https://cardio-ai.github.io/endovis-ml/ .

Marvin Seyfarth, Salman Ul Hassan Dar, Yannik Frisch et al.

Diffusion models have become a leading approach for high-fidelity medical image synthesis. However, most existing methods for 3D medical image generation rely on convolutional U-Net backbones within latent diffusion frameworks. While effective, these architectures impose strong locality biases and limited receptive fields, which may constrain scalability, global context integration, and flexible conditioning. In this work, we introduce VolDiT, the first purely transformer-based 3D Diffusion Transformer for volumetric medical image synthesis. Our approach extends diffusion transformers to native 3D data through volumetric patch embeddings and global self-attention operating directly over 3D tokens. To enable structured control, we propose a timestep-gated control adapter that maps segmentation masks into learnable control tokens that modulate transformer layers during denoising. This token-level conditioning mechanism allows precise spatial guidance while preserving the modeling advantages of transformer architectures. We evaluate our model on high-resolution 3D medical image synthesis tasks and compare it to state-of-the-art 3D latent diffusion models based on U-Nets. Results demonstrate improved global coherence, superior generative fidelity, and enhanced controllability. Our findings suggest that fully transformerbased diffusion models provide a flexible foundation for volumetric medical image synthesis. The code and models trained on public data are available at https://github.com/Cardio-AI/voldit.

Marvin Seyfarth, Salman Ul Hassan Dar, Sandy Engelhardt

Contemporary developments in generative AI are rapidly transforming the field of medical AI. These developments have been predominantly driven by the availability of large datasets and high computing power, which have facilitated a significant increase in model capacity. Despite their considerable potential, these models demand substantially high power, leading to high carbon dioxide (CO2) emissions. Given the harm such models are causing to the environment, there has been little focus on the carbon footprints of such models. This study analyzes carbon emissions from 2D and 3D latent diffusion models (LDMs) during training and data generation phases, revealing a surprising finding: the synthesis of large images contributes most significantly to these emissions. We assess different scenarios including model sizes, image dimensions, distributed training, and data generation steps. Our findings reveal substantial carbon emissions from these models, with training 2D and 3D models comparable to driving a car for 10 km and 90 km, respectively. The process of data generation is even more significant, with CO2 emissions equivalent to driving 160 km for 2D models and driving for up to 3345 km for 3D synthesis. Additionally, we found that the location of the experiment can increase carbon emissions by up to 94 times, and even the time of year can influence emissions by up to 50%. These figures are alarming, considering they represent only a single training and data generation phase for each model. Our results emphasize the urgent need for developing environmentally sustainable strategies in generative AI.

Sven Koehler, Sarah Kaye Mueller, Jonathan Kiekenap et al.

Cardiovascular magnetic resonance (CMR) is the gold standard for assessing cardiac function, but individual cardiac cycles complicate automatic temporal comparison or sub-phase analysis. Accurate cardiac keyframe detection can eliminate this problem. However, automatic methods solely derive end-systole (ES) and end-diastole (ED) frames from left ventricular volume curves, which do not provide a deeper insight into myocardial motion. We propose a self-supervised deep learning method detecting five keyframes in short-axis (SAX) and four-chamber long-axis (4CH) cine CMR. Initially, dense deformable registration fields are derived from the images and used to compute a 1D motion descriptor, which provides valuable insights into global cardiac contraction and relaxation patterns. From these characteristic curves, keyframes are determined using a simple set of rules. The method was independently evaluated for both views using three public, multicentre, multidisease datasets. M&Ms-2 (n=360) dataset was used for training and evaluation, and M&Ms (n=345) and ACDC (n=100) datasets for repeatability control. Furthermore, generalisability to patients with rare congenital heart defects was tested using the German Competence Network (GCN) dataset. Our self-supervised approach achieved improved detection accuracy by 30% - 51% for SAX and 11% - 47% for 4CH in ED and ES, as measured by cyclic frame difference (cFD), compared with the volume-based approach. We can detect ED and ES, as well as three additional keyframes throughout the cardiac cycle with a mean cFD below 1.31 frames for SAX and 1.73 for LAX. Our approach enables temporally aligned inter- and intra-patient analysis of cardiac dynamics, irrespective of cycle or phase lengths. GitHub repository: https://github.com/Cardio-AI/cmr-multi-view-phase-detection.git

Max-Heinrich Laves, Malte Tölle, Alexander Schlaefer et al.

We present Posterior Temperature Optimized Bayesian Inverse Models (POTOBIM), an unsupervised Bayesian approach to inverse problems in medical imaging using mean-field variational inference with a fully tempered posterior. Bayesian methods exhibit useful properties for approaching inverse tasks, such as tomographic reconstruction or image denoising. A suitable prior distribution introduces regularization, which is needed to solve the ill-posed problem and reduces overfitting the data. In practice, however, this often results in a suboptimal posterior temperature, and the full potential of the Bayesian approach is not being exploited. In POTOBIM, we optimize both the parameters of the prior distribution and the posterior temperature with respect to reconstruction accuracy using Bayesian optimization with Gaussian process regression. Our method is extensively evaluated on four different inverse tasks on a variety of modalities with images from public data sets and we demonstrate that an optimized posterior temperature outperforms both non-Bayesian and Bayesian approaches without temperature optimization. The use of an optimized prior distribution and posterior temperature leads to improved accuracy and uncertainty estimation and we show that it is sufficient to find these hyperparameters per task domain. Well-tempered posteriors yield calibrated uncertainty, which increases the reliability in the predictions. Our source code is publicly available at github.com/Cardio-AI/mfvi-dip-mia.

Sven Koehler, Lalith Sharan, Julian Kuhm et al.

Cardiac Magnetic Resonance (CMR) images are widely used for cardiac diagnosis and ventricular assessment. Extracting specific landmarks like the right ventricular insertion points is of importance for spatial alignment and 3D modeling. The automatic detection of such landmarks has been tackled by multiple groups using Deep Learning, but relatively little attention has been paid to the failure cases of evaluation metrics in this field. In this work, we extended the public ACDC dataset with additional labels of the right ventricular insertion points and compare different variants of a heatmap-based landmark detection pipeline. In this comparison, we demonstrate very likely pitfalls of apparently simple detection and localisation metrics which highlights the importance of a clear detection strategy and the definition of an upper limit for localisation-based metrics. Our preliminary results indicate that a combination of different metrics is necessary, as they yield different winners for method comparison. Additionally, they highlight the need of a comprehensive metric description and evaluation standardisation, especially for the error cases where no metrics could be computed or where no lower/upper boundary of a metric exists. Code and labels: https://github.com/Cardio-AI/rvip_landmark_detection

Lalith Sharan, Gabriele Romano, Julian Brand et al.

Purpose: Mitral valve repair is a complex minimally invasive surgery of the heart valve. In this context, suture detection from endoscopic images is a highly relevant task that provides quantitative information to analyse suturing patterns, assess prosthetic configurations and produce augmented reality visualisations. Facial or anatomical landmark detection tasks typically contain a fixed number of landmarks, and use regression or fixed heatmap-based approaches to localize the landmarks. However in endoscopy, there are a varying number of sutures in every image, and the sutures may occur at any location in the annulus, as they are not semantically unique. Method: In this work, we formulate the suture detection task as a multi-instance deep heatmap regression problem, to identify entry and exit points of sutures. We extend our previous work, and introduce the novel use of a 2D Gaussian layer followed by a differentiable 2D spatial Soft-Argmax layer to function as a local non-maximum suppression. Results: We present extensive experiments with multiple heatmap distribution functions and two variants of the proposed model. In the intra-operative domain, Variant 1 showed a mean F1 of +0.0422 over the baseline. Similarly, in the simulator domain, Variant 1 showed a mean F1 of +0.0865 over the baseline. Conclusion: The proposed model shows an improvement over the baseline in the intra-operative and the simulator domains. The data is made publicly available within the scope of the MICCAI AdaptOR2021 Challenge https://adaptor2021.github.io/, and the code at https://github.com/Cardio-AI/suture-detection-pytorch/. DOI:10.1007/s11548-021-02523-w. The link to the open access article can be found here: https://link.springer.com/article/10.1007%2Fs11548-021-02523-w

Lalith Sharan, Gabriele Romano, Sven Koehler et al.

The CycleGAN framework allows for unsupervised image-to-image translation of unpaired data. In a scenario of surgical training on a physical surgical simulator, this method can be used to transform endoscopic images of phantoms into images which more closely resemble the intra-operative appearance of the same surgical target structure. This can be viewed as a novel augmented reality approach, which we coined Hyperrealism in previous work. In this use case, it is of paramount importance to display objects like needles, sutures or instruments consistent in both domains while altering the style to a more tissue-like appearance. Segmentation of these objects would allow for a direct transfer, however, contouring of these, partly tiny and thin foreground objects is cumbersome and perhaps inaccurate. Instead, we propose to use landmark detection on the points when sutures pass into the tissue. This objective is directly incorporated into a CycleGAN framework by treating the performance of pre-trained detector models as an additional optimization goal. We show that a task defined on these sparse landmark labels improves consistency of synthesis by the generator network in both domains. Comparing a baseline CycleGAN architecture to our proposed extension (DetCycleGAN), mean precision (PPV) improved by +61.32, mean sensitivity (TPR) by +37.91, and mean F1 score by +0.4743. Furthermore, it could be shown that by dataset fusion, generated intra-operative images can be leveraged as additional training data for the detection network itself. The data is released within the scope of the AdaptOR MICCAI Challenge 2021 at https://adaptor2021.github.io/, and code at https://github.com/Cardio-AI/detcyclegan_pytorch.

Sven Koehler, Tarique Hussain, Zach Blair et al.

Anisotropic multi-slice Cardiac Magnetic Resonance (CMR) Images are conventionally acquired in patient-specific short-axis (SAX) orientation. In specific cardiovascular diseases that affect right ventricular (RV) morphology, acquisitions in standard axial (AX) orientation are preferred by some investigators, due to potential superiority in RV volume measurement for treatment planning. Unfortunately, due to the rare occurrence of these diseases, data in this domain is scarce. Recent research in deep learning-based methods mainly focused on SAX CMR images and they had proven to be very successful. In this work, we show that there is a considerable domain shift between AX and SAX images, and therefore, direct application of existing models yield sub-optimal results on AX samples. We propose a novel unsupervised domain adaptation approach, which uses task-related probabilities in an attention mechanism. Beyond that, cycle consistency is imposed on the learned patient-individual 3D rigid transformation to improve stability when automatically re-sampling the AX images to SAX orientations. The network was trained on 122 registered 3D AX-SAX CMR volume pairs from a multi-centric patient cohort. A mean 3D Dice of $0.86\pm{0.06}$ for the left ventricle, $0.65\pm{0.08}$ for the myocardium, and $0.77\pm{0.10}$ for the right ventricle could be achieved. This is an improvement of $25\%$ in Dice for RV in comparison to direct application on axial slices. To conclude, our pre-trained task module has neither seen CMR images nor labels from the target domain, but is able to segment them after the domain gap is reduced. Code: https://github.com/Cardio-AI/3d-mri-domain-adaptation

Salman Ul Hassan Dar, Marvin Seyfarth, Isabelle Ayx et al.

AI models present a wide range of applications in the field of medicine. However, achieving optimal performance requires access to extensive healthcare data, which is often not readily available. Furthermore, the imperative to preserve patient privacy restricts patient data sharing with third parties and even within institutes. Recently, generative AI models have been gaining traction for facilitating open-data sharing by proposing synthetic data as surrogates of real patient data. Despite the promise, some of these models are susceptible to patient data memorization, where models generate patient data copies instead of novel synthetic samples. Considering the importance of the problem, surprisingly it has received relatively little attention in the medical imaging community. To this end, we assess memorization in unconditional latent diffusion models. We train latent diffusion models on CT, MR, and X-ray datasets for synthetic data generation. We then detect the amount of training data memorized utilizing our novel self-supervised copy detection approach and further investigate various factors that can influence memorization. Our findings show a surprisingly high degree of patient data memorization across all datasets. Comparison with non-diffusion generative models, such as autoencoders and generative adversarial networks, indicates that while latent diffusion models are more susceptible to memorization, overall they outperform non-diffusion models in synthesis quality. Further analyses reveal that using augmentation strategies, small architecture, and increasing dataset can reduce memorization while over-training the models can enhance it. Collectively, our results emphasize the importance of carefully training generative models on private medical imaging datasets, and examining the synthetic data to ensure patient privacy before sharing it for medical research and applications.

Malte Tölle, Mohamad Scharaf, Samantha Fischer et al.

A patient undergoes multiple examinations in each hospital stay, where each provides different facets of the health status. These assessments include temporal data with varying sampling rates, discrete single-point measurements, therapeutic interventions such as medication administration, and images. While physicians are able to process and integrate diverse modalities intuitively, neural networks need specific modeling for each modality complicating the training procedure. We demonstrate that this complexity can be significantly reduced by visualizing all information as images along with unstructured text and subsequently training a conventional vision-text transformer. Our approach, Vision Transformer for irregular sampled Multi-modal Measurements (ViTiMM), not only simplifies data preprocessing and modeling but also outperforms current state-of-the-art methods in predicting in-hospital mortality and phenotyping, as evaluated on 6,175 patients from the MIMIC-IV dataset. The modalities include patient's clinical measurements, medications, X-ray images, and electrocardiography scans. We hope our work inspires advancements in multi-modal medical AI by reducing the training complexity to (visual) prompt engineering, thus lowering entry barriers and enabling no-code solutions for training. The source code will be made publicly available.

Marvin Seyfarth, Salman Ul Hassan Dar, Isabelle Ayx et al.

Advancements in AI for medical imaging offer significant potential. However, their applications are constrained by the limited availability of data and the reluctance of medical centers to share it due to patient privacy concerns. Generative models present a promising solution by creating synthetic data as a substitute for real patient data. However, medical images are typically high-dimensional, and current state-of-the-art methods are often impractical for computational resource-constrained healthcare environments. These models rely on data sub-sampling, raising doubts about their feasibility and real-world applicability. Furthermore, many of these models are evaluated on quantitative metrics that alone can be misleading in assessing the image quality and clinical meaningfulness of the generated images. To address this, we introduce MedLoRD, a generative diffusion model designed for computational resource-constrained environments. MedLoRD is capable of generating high-dimensional medical volumes with resolutions up to 512$\times$512$\times$256, utilizing GPUs with only 24GB VRAM, which are commonly found in standard desktop workstations. MedLoRD is evaluated across multiple modalities, including Coronary Computed Tomography Angiography and Lung Computed Tomography datasets. Extensive evaluations through radiological evaluation, relative regional volume analysis, adherence to conditional masks, and downstream tasks show that MedLoRD generates high-fidelity images closely adhering to segmentation mask conditions, surpassing the capabilities of current state-of-the-art generative models for medical image synthesis in computational resource-constrained environments.

Lukas Burger, Lalith Sharan, Samantha Fischer et al.

To support minimally-invasive intraoperative mitral valve repair, quantitative measurements from the valve can be obtained using an infra-red tracked stylus. It is desirable to view such manually measured points together with the endoscopic image for further assistance. Therefore, hand-eye calibration is required that links both coordinate systems and is a prerequisite to project the points onto the image plane. A complementary approach to this is to use a vision-based endoscopic stereo-setup to detect and triangulate points of interest, to obtain the 3D coordinates. In this paper, we aim to compare both approaches on a rigid phantom and two patient-individual silicone replica which resemble the intraoperative scenario. The preliminary results indicate that 3D landmark estimation, either labeled manually or through partly automated detection with a deep learning approach, provides more accurate triangulated depth measurements when performed with a tailored image-based method than with stylus measurements.

Max-Heinrich Laves, Malte Tölle, Alexander Schlaefer et al.

Bayesian methods feature useful properties for solving inverse problems, such as tomographic reconstruction. The prior distribution introduces regularization, which helps solving the ill-posed problem and reduces overfitting. In practice, this often results in a suboptimal posterior temperature and the full potential of the Bayesian approach is not realized. In this paper, we optimize both the parameters of the prior distribution and the posterior temperature using Bayesian optimization. Well-tempered posteriors lead to better predictive performance and improved uncertainty calibration, which we demonstrate for the task of sparse-view CT reconstruction.

Annika Reinke, Minu D. Tizabi, Carole H. Sudre et al.

While the importance of automatic image analysis is continuously increasing, recent meta-research revealed major flaws with respect to algorithm validation. Performance metrics are particularly key for meaningful, objective, and transparent performance assessment and validation of the used automatic algorithms, but relatively little attention has been given to the practical pitfalls when using specific metrics for a given image analysis task. These are typically related to (1) the disregard of inherent metric properties, such as the behaviour in the presence of class imbalance or small target structures, (2) the disregard of inherent data set properties, such as the non-independence of the test cases, and (3) the disregard of the actual biomedical domain interest that the metrics should reflect. This living dynamically document has the purpose to illustrate important limitations of performance metrics commonly applied in the field of image analysis. In this context, it focuses on biomedical image analysis problems that can be phrased as image-level classification, semantic segmentation, instance segmentation, or object detection task. The current version is based on a Delphi process on metrics conducted by an international consortium of image analysis experts from more than 60 institutions worldwide.

Antonia Stern, Lalith Sharan, Gabriele Romano et al.

Mitral valve repair is a surgery to restore the function of the mitral valve. To achieve this, a prosthetic ring is sewed onto the mitral annulus. Analyzing the sutures, which are punctured through the annulus for ring implantation, can be useful in surgical skill assessment, for quantitative surgery and for positioning a virtual prosthetic ring model in the scene via augmented reality. This work presents a neural network approach which detects the sutures in endoscopic images of mitral valve repair and therefore solves a landmark detection problem with varying amount of landmarks, as opposed to most other existing deep learning-based landmark detection approaches. The neural network is trained separately on two data collections from different domains with the same architecture and hyperparameter settings. The datasets consist of more than 1,300 stereo frame pairs each, with a total over 60,000 annotated landmarks. The proposed heatmap-based neural network achieves a mean positive predictive value (PPV) of 66.68$\pm$4.67% and a mean true positive rate (TPR) of 24.45$\pm$5.06% on the intraoperative test dataset and a mean PPV of 81.50\pm5.77\% and a mean TPR of 61.60$\pm$6.11% on a dataset recorded during surgical simulation. The best detection results are achieved when the camera is positioned above the mitral valve with good illumination. A detection from a sideward view is also possible if the mitral valve is well perceptible.

Zhaohan Xiong, Qing Xia, Zhiqiang Hu et al.

Segmentation of cardiac images, particularly late gadolinium-enhanced magnetic resonance imaging (LGE-MRI) widely used for visualizing diseased cardiac structures, is a crucial first step for clinical diagnosis and treatment. However, direct segmentation of LGE-MRIs is challenging due to its attenuated contrast. Since most clinical studies have relied on manual and labor-intensive approaches, automatic methods are of high interest, particularly optimized machine learning approaches. To address this, we organized the "2018 Left Atrium Segmentation Challenge" using 154 3D LGE-MRIs, currently the world's largest cardiac LGE-MRI dataset, and associated labels of the left atrium segmented by three medical experts, ultimately attracting the participation of 27 international teams. In this paper, extensive analysis of the submitted algorithms using technical and biological metrics was performed by undergoing subgroup analysis and conducting hyper-parameter analysis, offering an overall picture of the major design choices of convolutional neural networks (CNNs) and practical considerations for achieving state-of-the-art left atrium segmentation. Results show the top method achieved a dice score of 93.2% and a mean surface to a surface distance of 0.7 mm, significantly outperforming prior state-of-the-art. Particularly, our analysis demonstrated that double, sequentially used CNNs, in which a first CNN is used for automatic region-of-interest localization and a subsequent CNN is used for refined regional segmentation, achieved far superior results than traditional methods and pipelines containing single CNNs. This large-scale benchmarking study makes a significant step towards much-improved segmentation methods for cardiac LGE-MRIs, and will serve as an important benchmark for evaluating and comparing the future works in the field.

Sven Koehler, Animesh Tandon, Tarique Hussain et al.

Planning the optimal time of intervention for pulmonary valve replacement surgery in patients with the congenital heart disease Tetralogy of Fallot (TOF) is mainly based on ventricular volume and function according to current guidelines. Both of these two biomarkers are most reliably assessed by segmentation of 3D cardiac magnetic resonance (CMR) images. In several grand challenges in the last years, U-Net architectures have shown impressive results on the provided data. However, in clinical practice, data sets are more diverse considering individual pathologies and image properties derived from different scanner properties. Additionally, specific training data for complex rare diseases like TOF is scarce. For this work, 1) we assessed the accuracy gap when using a publicly available labelled data set (the Automatic Cardiac Diagnosis Challenge (ACDC) data set) for training and subsequent applying it to CMR data of TOF patients and vice versa and 2) whether we can achieve similar results when applying the model to a more heterogeneous data base. Multiple deep learning models were trained with four-fold cross validation. Afterwards they were evaluated on the respective unseen CMR images from the other collection. Our results confirm that current deep learning models can achieve excellent results (left ventricle dice of $0.951\pm{0.003}$/$0.941\pm{0.007}$ train/validation) within a single data collection. But once they are applied to other pathologies, it becomes apparent how much they overfit to the training pathologies (dice score drops between $0.072\pm{0.001}$ for the left and $0.165\pm{0.001}$ for the right ventricle).

Chandrakanth Jayachandran Preetha, Jonathan Kloss, Fabian Siegfried Wehrtmann et al.

Minimally Invasive Surgery (MIS) techniques have gained rapid popularity among surgeons since they offer significant clinical benefits including reduced recovery time and diminished post-operative adverse effects. However, conventional endoscopic systems output monocular video which compromises depth perception, spatial orientation and field of view. Suturing is one of the most complex tasks performed under these circumstances. Key components of this tasks are the interplay between needle holder and the surgical needle. Reliable 3D localization of needle and instruments in real time could be used to augment the scene with additional parameters that describe their quantitative geometric relation, e.g. the relation between the estimated needle plane and its rotation center and the instrument. This could contribute towards standardization and training of basic skills and operative techniques, enhance overall surgical performance, and reduce the risk of complications. The paper proposes an Augmented Reality environment with quantitative and qualitative visual representations to enhance laparoscopic training outcomes performed on a silicone pad. This is enabled by a multi-task supervised deep neural network which performs multi-class segmentation and depth map prediction. Scarcity of labels has been conquered by creating a virtual environment which resembles the surgical training scenario to generate dense depth maps and segmentation maps. The proposed convolutional neural network was tested on real surgical training scenarios and showed to be robust to occlusion of the needle. The network achieves a dice score of 0.67 for surgical needle segmentation, 0.81 for needle holder instrument segmentation and a mean absolute error of 6.5 mm for depth estimation.

Micha Pfeiffer, Isabel Funke, Maria R. Robu et al.

In the medical domain, the lack of large training data sets and benchmarks is often a limiting factor for training deep neural networks. In contrast to expensive manual labeling, computer simulations can generate large and fully labeled data sets with a minimum of manual effort. However, models that are trained on simulated data usually do not translate well to real scenarios. To bridge the domain gap between simulated and real laparoscopic images, we exploit recent advances in unpaired image-to-image translation. We extent an image-to-image translation method to generate a diverse multitude of realistically looking synthetic images based on images from a simple laparoscopy simulation. By incorporating means to ensure that the image content is preserved during the translation process, we ensure that the labels given for the simulated images remain valid for their realistically looking translations. This way, we are able to generate a large, fully labeled synthetic data set of laparoscopic images with realistic appearance. We show that this data set can be used to train models for the task of liver segmentation of laparoscopic images. We achieve average dice scores of up to 0.89 in some patients without manually labeling a single laparoscopic image and show that using our synthetic data to pre-train models can greatly improve their performance. The synthetic data set will be made publicly available, fully labeled with segmentation maps, depth maps, normal maps, and positions of tools and camera (http://opencas.dkfz.de/image2image).

Sandy Engelhardt, Lalith Sharan, Matthias Karck et al.

Phantoms for surgical training are able to mimic cutting and suturing properties and patient-individual shape of organs, but lack a realistic visual appearance that captures the heterogeneity of surgical scenes. In order to overcome this in endoscopic approaches, hyperrealistic concepts have been proposed to be used in an augmented reality-setting, which are based on deep image-to-image transformation methods. Such concepts are able to generate realistic representations of phantoms learned from real intraoperative endoscopic sequences. Conditioned on frames from the surgical training process, the learned models are able to generate impressive results by transforming unrealistic parts of the image (e.g.\ the uniform phantom texture is replaced by the more heterogeneous texture of the tissue). Image-to-image synthesis usually learns a mapping $G:X~\to~Y$ such that the distribution of images from $G(X)$ is indistinguishable from the distribution $Y$. However, it does not necessarily force the generated images to be consistent and without artifacts. In the endoscopic image domain this can affect depth cues and stereo consistency of a stereo image pair, which ultimately impairs surgical vision. We propose a cross-domain conditional generative adversarial network approach (GAN) that aims to generate more consistent stereo pairs. The results show substantial improvements in depth perception and realism evaluated by 3 domain experts and 3 medical students on a 3D monitor over the baseline method. In 84 of 90 instances our proposed method was preferred or rated equal to the baseline.

Fabian Isensee, Paul Jaeger, Peter M. Full et al.

Cardiac magnetic resonance imaging improves on diagnosis of cardiovascular diseases by providing images at high spatiotemporal resolution. Manual evaluation of these time-series, however, is expensive and prone to biased and non-reproducible outcomes. In this paper, we present a method that addresses named limitations by integrating segmentation and disease classification into a fully automatic processing pipeline. We use an ensemble of UNet inspired architectures for segmentation of cardiac structures such as the left and right ventricular cavity (LVC, RVC) and the left ventricular myocardium (LVM) on each time instance of the cardiac cycle. For the classification task, information is extracted from the segmented time-series in form of comprehensive features handcrafted to reflect diagnostic clinical procedures. Based on these features we train an ensemble of heavily regularized multilayer perceptrons (MLP) and a random forest classifier to predict the pathologic target class. We evaluated our method on the ACDC dataset (4 pathology groups, 1 healthy group) and achieve dice scores of 0.945 (LVC), 0.908 (RVC) and 0.905 (LVM) in a cross-validation over the training set (100 cases) and 0.950 (LVC), 0.923 (RVC) and 0.911 (LVM) on the test set (50 cases). We report a classification accuracy of 94% on a training set cross-validation and 92% on the test set. Our results underpin the potential of machine learning methods for accurate, fast and reproducible segmentation and computer-assisted diagnosis (CAD).