David Zimmerer

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.

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.

Saikat Roy, Tassilo Wald, Gregor Koehler et al.

Foundation models have taken over natural language processing and image generation domains due to the flexibility of prompting. With the recent introduction of the Segment Anything Model (SAM), this prompt-driven paradigm has entered image segmentation with a hitherto unexplored abundance of capabilities. The purpose of this paper is to conduct an initial evaluation of the out-of-the-box zero-shot capabilities of SAM for medical image segmentation, by evaluating its performance on an abdominal CT organ segmentation task, via point or bounding box based prompting. We show that SAM generalizes well to CT data, making it a potential catalyst for the advancement of semi-automatic segmentation tools for clinicians. We believe that this foundation model, while not reaching state-of-the-art segmentation performance in our investigations, can serve as a highly potent starting point for further adaptations of such models to the intricacies of the medical domain. Keywords: medical image segmentation, SAM, foundation models, zero-shot learning

Stefan Denner, Markus Bujotzek, Dimitrios Bounias et al.

Medical image classification plays a crucial role in clinical decision-making, yet most models are constrained to a fixed set of predefined classes, limiting their adaptability to new conditions. Contrastive Language-Image Pretraining (CLIP) offers a promising solution by enabling zero-shot classification through multimodal large-scale pretraining. However, while CLIP effectively captures global image content, radiology requires a more localized focus on specific pathology regions to enhance both interpretability and diagnostic accuracy. To address this, we explore the potential of incorporating visual cues into zero-shot classification, embedding visual markers, such as arrows, bounding boxes, and circles, directly into radiological images to guide model attention. Evaluating across four public chest X-ray datasets, we demonstrate that visual markers improve AUROC by up to 0.185, highlighting their effectiveness in enhancing classification performance. Furthermore, attention map analysis confirms that visual cues help models focus on clinically relevant areas, leading to more interpretable predictions.To support further research, we use public datasets and provide our codebase and preprocessing pipeline under https://github.com/MIC-DKFZ/VPE-in-Radiology, serving as a reference point for future work on localized classification in medical imaging.

Carsten T. Lüth, David Zimmerer, Gregor Koehler et al.

Unsupervised anomaly detection in medical imaging aims to detect and localize arbitrary anomalies without requiring annotated anomalous data during training. Often, this is achieved by learning a data distribution of normal samples and detecting anomalies as regions in the image which deviate from this distribution. Most current state-of-the-art methods use latent variable generative models operating directly on the images. However, generative models have been shown to mostly capture low-level features, s.a. pixel-intensities, instead of rich semantic features, which also applies to their representations. We circumvent this problem by proposing CRADL whose core idea is to model the distribution of normal samples directly in the low-dimensional representation space of an encoder trained with a contrastive pretext-task. By utilizing the representations of contrastive learning, we aim to fix the over-fixation on low-level features and learn more semantic-rich representations. Our experiments on anomaly detection and localization tasks using three distinct evaluation datasets show that 1) contrastive representations are superior to representations of generative latent variable models and 2) the CRADL framework shows competitive or superior performance to state-of-the-art.

Gregor Koehler, Tassilo Wald, Constantin Ulrich et al.

Despite the remarkable success of deep learning systems over the last decade, a key difference still remains between neural network and human decision-making: As humans, we cannot only form a decision on the spot, but also ponder, revisiting an initial guess from different angles, distilling relevant information, arriving at a better decision. Here, we propose RecycleNet, a latent feature recycling method, instilling the pondering capability for neural networks to refine initial decisions over a number of recycling steps, where outputs are fed back into earlier network layers in an iterative fashion. This approach makes minimal assumptions about the neural network architecture and thus can be implemented in a wide variety of contexts. Using medical image segmentation as the evaluation environment, we show that latent feature recycling enables the network to iteratively refine initial predictions even beyond the iterations seen during training, converging towards an improved decision. We evaluate this across a variety of segmentation benchmarks and show consistent improvements even compared with top-performing segmentation methods. This allows trading increased computation time for improved performance, which can be beneficial, especially for safety-critical applications.

Tassilo Wald, Constantin Ulrich, Fabian Isensee et al.

Independently trained machine learning models tend to learn similar features. Given an ensemble of independently trained models, this results in correlated predictions and common failure modes. Previous attempts focusing on decorrelation of output predictions or logits yielded mixed results, particularly due to their reduction in model accuracy caused by conflicting optimization objectives. In this paper, we propose the novel idea of utilizing methods of the representational similarity field to promote dissimilarity during training instead of measuring similarity of trained models. To this end, we promote intermediate representations to be dissimilar at different depths between architectures, with the goal of learning robust ensembles with disjoint failure modes. We show that highly dissimilar intermediate representations result in less correlated output predictions and slightly lower error consistency, resulting in higher ensemble accuracy. With this, we shine first light on the connection between intermediate representations and their impact on the output predictions.

Nico Albert Disch, Saikat Roy, Constantin Ulrich et al.

Forecasting how 3D medical scans evolve over time is important for disease progression, treatment planning, and developmental assessment. Yet existing models either rely on a single prior scan, fixed grid times, or target global labels, which limits voxel-level forecasting under irregular sampling. We present CRONOS, a unified framework for many-to-one prediction from multiple past scans that supports both discrete (grid-based) and continuous (real-valued) timestamps in one model, to the best of our knowledge the first to achieve continuous sequence-to-image forecasting for 3D medical data. CRONOS learns a spatio-temporal velocity field that transports context volumes toward a target volume at an arbitrary time, while operating directly in 3D voxel space. Across three public datasets spanning Cine-MRI, perfusion CT, and longitudinal MRI, CRONOS outperforms other baselines, while remaining computationally competitive. We will release code and evaluation protocols to enable reproducible, multi-dataset benchmarking of multi-context, continuous-time forecasting.

Stefan Denner, David Zimmerer, Dimitrios Bounias et al.

Content-based image retrieval (CBIR) has the potential to significantly improve diagnostic aid and medical research in radiology. However, current CBIR systems face limitations due to their specialization to certain pathologies, limiting their utility. On the other hand, several vision foundation models have been shown to produce general-purpose visual features. Therefore, in this work, we propose using vision foundation models as powerful and versatile off-the-shelf feature extractors for content-based image retrieval. Our contributions include: (1) benchmarking a diverse set of vision foundation models on an extensive dataset comprising 1.6 million 2D radiological images across four modalities and 161 pathologies; (2) identifying weakly-supervised models, particularly BiomedCLIP, as highly effective, achieving a achieving a P@1 of up to 0.594 (P@3: 0.590, P@5: 0.588, P@10: 0.583), comparable to specialized CBIR systems but without additional training; (3) conducting an in-depth analysis of the impact of index size on retrieval performance; (4) evaluating the quality of embedding spaces generated by different models; and (5) investigating specific challenges associated with retrieving anatomical versus pathological structures. Despite these challenges, our research underscores the vast potential of foundation models for CBIR in radiology, proposing a shift towards versatile, general-purpose medical image retrieval systems that do not require specific tuning. Our code, dataset splits and embeddings are publicly available under https://github.com/MIC-DKFZ/foundation-models-for-cbmir.

Alexandra Ertl, Shuhan Xiao, Stefan Denner et al.

Landmark detection plays a crucial role in medical imaging tasks that rely on precise spatial localization, including specific applications in diagnosis, treatment planning, image registration, and surgical navigation. However, manual annotation is labor-intensive and requires expert knowledge. While deep learning shows promise in automating this task, progress is hindered by limited public datasets, inconsistent benchmarks, and non-standardized baselines, restricting reproducibility, fair comparisons, and model generalizability. This work introduces nnLandmark, a self-configuring deep learning framework for 3D medical landmark detection, adapting nnU-Net to perform heatmap-based regression. By leveraging nnU-Net's automated configuration, nnLandmark eliminates the need for manual parameter tuning, offering out-of-the-box usability. It achieves state-of-the-art accuracy across two public datasets, with a mean radial error (MRE) of 1.5 mm on the Mandibular Molar Landmark (MML) dental CT dataset and 1.2 mm for anatomical fiducials on a brain MRI dataset (AFIDs), where nnLandmark aligns with the inter-rater variability of 1.5 mm. With its strong generalization, reproducibility, and ease of deployment, nnLandmark establishes a reliable baseline for 3D landmark detection, supporting research in anatomical localization and clinical workflows that depend on precise landmark identification. The code will be available soon.

Nico Albert Disch, Yannick Kirchhoff, Robin Peretzke et al.

Understanding temporal dynamics in medical imaging is crucial for applications such as disease progression modeling, treatment planning and anatomical development tracking. However, most deep learning methods either consider only single temporal contexts, or focus on tasks like classification or regression, limiting their ability for fine-grained spatial predictions. While some approaches have been explored, they are often limited to single timepoints, specific diseases or have other technical restrictions. To address this fundamental gap, we introduce Temporal Flow Matching (TFM), a unified generative trajectory method that (i) aims to learn the underlying temporal distribution, (ii) by design can fall back to a nearest image predictor, i.e. predicting the last context image (LCI), as a special case, and (iii) supports $3D$ volumes, multiple prior scans, and irregular sampling. Extensive benchmarks on three public longitudinal datasets show that TFM consistently surpasses spatio-temporal methods from natural imaging, establishing a new state-of-the-art and robust baseline for $4D$ medical image prediction.

Tassilo Wald, Constantin Ulrich, Gregor Köhler et al.

What representation do deep neural networks learn? How similar are images to each other for neural networks? Despite the overwhelming success of deep learning methods key questions about their internal workings still remain largely unanswered, due to their internal high dimensionality and complexity. To address this, one approach is to measure the similarity of activation responses to various inputs. Representational Similarity Matrices (RSMs) distill this similarity into scalar values for each input pair. These matrices encapsulate the entire similarity structure of a system, indicating which input leads to similar responses. While the similarity between images is ambiguous, we argue that the spatial location of semantic objects does neither influence human perception nor deep learning classifiers. Thus this should be reflected in the definition of similarity between image responses for computer vision systems. Revisiting the established similarity calculations for RSMs we expose their sensitivity to spatial alignment. In this paper, we propose to solve this through semantic RSMs, which are invariant to spatial permutation. We measure semantic similarity between input responses by formulating it as a set-matching problem. Further, we quantify the superiority of semantic RSMs over spatio-semantic RSMs through image retrieval and by comparing the similarity between representations to the similarity between predicted class probabilities.

Maximilian Zenk, David Zimmerer, Fabian Isensee et al.

Semantic segmentation is an essential component of medical image analysis research, with recent deep learning algorithms offering out-of-the-box applicability across diverse datasets. Despite these advancements, segmentation failures remain a significant concern for real-world clinical applications, necessitating reliable detection mechanisms. This paper introduces a comprehensive benchmarking framework aimed at evaluating failure detection methodologies within medical image segmentation. Through our analysis, we identify the strengths and limitations of current failure detection metrics, advocating for the risk-coverage analysis as a holistic evaluation approach. Utilizing a collective dataset comprising five public 3D medical image collections, we assess the efficacy of various failure detection strategies under realistic test-time distribution shifts. Our findings highlight the importance of pixel confidence aggregation and we observe superior performance of the pairwise Dice score (Roy et al., 2019) between ensemble predictions, positioning it as a simple and robust baseline for failure detection in medical image segmentation. To promote ongoing research, we make the benchmarking framework available to the community.

Jens Petersen, Fabian Isensee, Gregor Köhler et al.

The ability to estimate how a tumor might evolve in the future could have tremendous clinical benefits, from improved treatment decisions to better dose distribution in radiation therapy. Recent work has approached the glioma growth modeling problem via deep learning and variational inference, thus learning growth dynamics entirely from a real patient data distribution. So far, this approach was constrained to predefined image acquisition intervals and sequences of fixed length, which limits its applicability in more realistic scenarios. We overcome these limitations by extending Neural Processes, a class of conditional generative models for stochastic time series, with a hierarchical multi-scale representation encoding including a spatio-temporal attention mechanism. The result is a learned growth model that can be conditioned on an arbitrary number of observations, and that can produce a distribution of temporally consistent growth trajectories on a continuous time axis. On a dataset of 379 patients, the approach successfully captures both global and finer-grained variations in the images, exhibiting superior performance compared to other learned growth models.

Jens Petersen, Gregor Köhler, David Zimmerer et al.

Neural Processes (NPs) are a family of conditional generative models that are able to model a distribution over functions, in a way that allows them to perform predictions at test time conditioned on a number of context points. A recent addition to this family, Convolutional Conditional Neural Processes (ConvCNP), have shown remarkable improvement in performance over prior art, but we find that they sometimes struggle to generalize when applied to time series data. In particular, they are not robust to distribution shifts and fail to extrapolate observed patterns into the future. By incorporating a Gaussian Process into the model, we are able to remedy this and at the same time improve performance within distribution. As an added benefit, the Gaussian Process reintroduces the possibility to sample from the model, a key feature of other members in the NP family.

Sarthak Pati, Ujjwal Baid, Maximilian Zenk et al.

This manuscript describes the first challenge on Federated Learning, namely the Federated Tumor Segmentation (FeTS) challenge 2021. International challenges have become the standard for validation of biomedical image analysis methods. However, the actual performance of participating (even the winning) algorithms on "real-world" clinical data often remains unclear, as the data included in challenges are usually acquired in very controlled settings at few institutions. The seemingly obvious solution of just collecting increasingly more data from more institutions in such challenges does not scale well due to privacy and ownership hurdles. Towards alleviating these concerns, we are proposing the FeTS challenge 2021 to cater towards both the development and the evaluation of models for the segmentation of intrinsically heterogeneous (in appearance, shape, and histology) brain tumors, namely gliomas. Specifically, the FeTS 2021 challenge uses clinically acquired, multi-institutional magnetic resonance imaging (MRI) scans from the BraTS 2020 challenge, as well as from various remote independent institutions included in the collaborative network of a real-world federation (https://www.fets.ai/). The goals of the FeTS challenge are directly represented by the two included tasks: 1) the identification of the optimal weight aggregation approach towards the training of a consensus model that has gained knowledge via federated learning from multiple geographically distinct institutions, while their data are always retained within each institution, and 2) the federated evaluation of the generalizability of brain tumor segmentation models "in the wild", i.e. on data from institutional distributions that were not part of the training datasets.

David Zimmerer, Jens Petersen, Simon A. A. Kohl et al.

Through training on unlabeled data, anomaly detection has the potential to impact computer-aided diagnosis by outlining suspicious regions. Previous work on deep-learning-based anomaly detection has primarily focused on the reconstruction error. We argue instead, that pixel-wise anomaly ratings derived from a Variational Autoencoder based score approximation yield a theoretically better grounded and more faithful estimate. In our experiments, Variational Autoencoder gradient-based rating outperforms other approaches on unsupervised pixel-wise tumor detection on the BraTS-2017 dataset with a ROC-AUC of 0.94.

David Zimmerer, Jens Petersen, Klaus Maier-Hein

Variational Auto-Encoders have often been used for unsupervised pretraining, feature extraction and out-of-distribution and anomaly detection in the medical field. However, VAEs often lack the ability to produce sharp images and learn high-level features. We propose to alleviate these issues by adding a new branch to conditional hierarchical VAEs. This enforces a division between higher-level and lower-level features. Despite the additional computational overhead compared to a normal VAE it results in sharper and better reconstructions and can capture the data distribution similarly well (indicated by a similar or slightly better OoD detection performance).

David Zimmerer, Fabian Isensee, Jens Petersen et al.

An assumption-free automatic check of medical images for potentially overseen anomalies would be a valuable assistance for a radiologist. Deep learning and especially Variational Auto-Encoders (VAEs) have shown great potential in the unsupervised learning of data distributions. In principle, this allows for such a check and even the localization of parts in the image that are most suspicious. Currently, however, the reconstruction-based localization by design requires adjusting the model architecture to the specific problem looked at during evaluation. This contradicts the principle of building assumption-free models. We propose complementing the localization part with a term derived from the Kullback-Leibler (KL)-divergence. For validation, we perform a series of experiments on FashionMNIST as well as on a medical task including >1000 healthy and >250 brain tumor patients. Results show that the proposed formalism outperforms the state of the art VAE-based localization of anomalies across many hyperparameter settings and also shows a competitive max performance.

David Zimmerer, Simon A. A. Kohl, Jens Petersen et al.

Unsupervised learning can leverage large-scale data sources without the need for annotations. In this context, deep learning-based auto encoders have shown great potential in detecting anomalies in medical images. However, state-of-the-art anomaly scores are still based on the reconstruction error, which lacks in two essential parts: it ignores the model-internal representation employed for reconstruction, and it lacks formal assertions and comparability between samples. We address these shortcomings by proposing the Context-encoding Variational Autoencoder (ceVAE) which combines reconstruction- with density-based anomaly scoring. This improves the sample- as well as pixel-wise results. In our experiments on the BraTS-2017 and ISLES-2015 segmentation benchmarks, the ceVAE achieves unsupervised ROC-AUCs of 0.95 and 0.89, respectively, thus outperforming state-of-the-art methods by a considerable margin.

Fabian Isensee, Jens Petersen, Andre Klein et al.

The U-Net was presented in 2015. With its straight-forward and successful architecture it quickly evolved to a commonly used benchmark in medical image segmentation. The adaptation of the U-Net to novel problems, however, comprises several degrees of freedom regarding the exact architecture, preprocessing, training and inference. These choices are not independent of each other and substantially impact the overall performance. The present paper introduces the nnU-Net ('no-new-Net'), which refers to a robust and self-adapting framework on the basis of 2D and 3D vanilla U-Nets. We argue the strong case for taking away superfluous bells and whistles of many proposed network designs and instead focus on the remaining aspects that make out the performance and generalizability of a method. We evaluate the nnU-Net in the context of the Medical Segmentation Decathlon challenge, which measures segmentation performance in ten disciplines comprising distinct entities, image modalities, image geometries and dataset sizes, with no manual adjustments between datasets allowed. At the time of manuscript submission, nnU-Net achieves the highest mean dice scores across all classes and seven phase 1 tasks (except class 1 in BrainTumour) in the online leaderboard of the challenge.

Tobias Ross, David Zimmerer, Anant Vemuri et al.

Surgical data science is a new research field that aims to observe all aspects of the patient treatment process in order to provide the right assistance at the right time. Due to the breakthrough successes of deep learning-based solutions for automatic image annotation, the availability of reference annotations for algorithm training is becoming a major bottleneck in the field. The purpose of this paper was to investigate the concept of self-supervised learning to address this issue. Our approach is guided by the hypothesis that unlabeled video data can be used to learn a representation of the target domain that boosts the performance of state-of-the-art machine learning algorithms when used for pre-training. Core of the method is an auxiliary task based on raw endoscopic video data of the target domain that is used to initialize the convolutional neural network (CNN) for the target task. In this paper, we propose the re-colorization of medical images with a generative adversarial network (GAN)-based architecture as auxiliary task. A variant of the method involves a second pre-training step based on labeled data for the target task from a related domain. We validate both variants using medical instrument segmentation as target task. The proposed approach can be used to radically reduce the manual annotation effort involved in training CNNs. Compared to the baseline approach of generating annotated data from scratch, our method decreases exploratively the number of labeled images by up to 75% without sacrificing performance. Our method also outperforms alternative methods for CNN pre-training, such as pre-training on publicly available non-medical or medical data using the target task (in this instance: segmentation). As it makes efficient use of available (non-)public and (un-)labeled data, the approach has the potential to become a valuable tool for CNN (pre-)training.