Touchstone Benchmark: Are We on the Right Way for Evaluating AI Algorithms for Medical Segmentation?

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.

Decoupling Semantic Similarity from Spatial Alignment for Neural Networks

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.

Data-Centric Strategies for Overcoming PET/CT Heterogeneity: Insights from the AutoPET III Lesion Segmentation Challenge

The third autoPET challenge introduced a new data-centric task this year, shifting the focus from model development to improving metastatic lesion segmentation on PET/CT images through data quality and handling strategies. In response, we developed targeted methods to enhance segmentation performance tailored to the characteristics of PET/CT imaging. Our approach encompasses two key elements. First, to address potential alignment errors between CT and PET modalities as well as the prevalence of punctate lesions, we modified the baseline data augmentation scheme and extended it with misalignment augmentation. This adaptation aims to improve segmentation accuracy, particularly for tiny metastatic lesions. Second, to tackle the variability in image dimensions significantly affecting the prediction time, we implemented a dynamic ensembling and test-time augmentation (TTA) strategy. This method optimizes the use of ensembling and TTA within a 5-minute prediction time limit, effectively leveraging the generalization potential for both small and large images. Both of our solutions are designed to be robust across different tracers and institutional settings, offering a general, yet imaging-specific approach to the multi-tracer and multi-institutional challenges of the competition. We made the challenge repository with our modifications publicly available at \url{https://github.com/MIC-DKFZ/miccai2024_autopet3_datacentric}.

From FDG to PSMA: A Hitchhiker's Guide to Multitracer, Multicenter Lesion Segmentation in PET/CT Imaging

Automated lesion segmentation in PET/CT scans is crucial for improving clinical workflows and advancing cancer diagnostics. However, the task is challenging due to physiological variability, different tracers used in PET imaging, and diverse imaging protocols across medical centers. To address this, the autoPET series was created to challenge researchers to develop algorithms that generalize across diverse PET/CT environments. This paper presents our solution for the autoPET III challenge, targeting multitracer, multicenter generalization using the nnU-Net framework with the ResEncL architecture. Key techniques include misalignment data augmentation and multi-modal pretraining across CT, MR, and PET datasets to provide an initial anatomical understanding. We incorporate organ supervision as a multitask approach, enabling the model to distinguish between physiological uptake and tracer-specific patterns, which is particularly beneficial in cases where no lesions are present. Compared to the default nnU-Net, which achieved a Dice score of 57.61, or the larger ResEncL (65.31) our model significantly improved performance with a Dice score of 68.40, alongside a reduction in false positive (FPvol: 7.82) and false negative (FNvol: 10.35) volumes. These results underscore the effectiveness of combining advanced network design, augmentation, pretraining, and multitask learning for PET/CT lesion segmentation. Code is publicly available at https://github.com/MIC-DKFZ/autopet-3-submission.

Quality Assured: Rethinking Annotation Strategies in Imaging AI

This paper does not describe a novel method. Instead, it studies an essential foundation for reliable benchmarking and ultimately real-world application of AI-based image analysis: generating high-quality reference annotations. Previous research has focused on crowdsourcing as a means of outsourcing annotations. However, little attention has so far been given to annotation companies, specifically regarding their internal quality assurance (QA) processes. Therefore, our aim is to evaluate the influence of QA employed by annotation companies on annotation quality and devise methodologies for maximizing data annotation efficacy. Based on a total of 57,648 instance segmented images obtained from a total of 924 annotators and 34 QA workers from four annotation companies and Amazon Mechanical Turk (MTurk), we derived the following insights: (1) Annotation companies perform better both in terms of quantity and quality compared to the widely used platform MTurk. (2) Annotation companies' internal QA only provides marginal improvements, if any. However, improving labeling instructions instead of investing in QA can substantially boost annotation performance. (3) The benefit of internal QA depends on specific image characteristics. Our work could enable researchers to derive substantially more value from a fixed annotation budget and change the way annotation companies conduct internal QA.

Comparative Benchmarking of Failure Detection Methods in Medical Image Segmentation: Unveiling the Role of Confidence Aggregation

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.

Mitigating False Predictions In Unreasonable Body Regions

Despite considerable strides in developing deep learning models for 3D medical image segmentation, the challenge of effectively generalizing across diverse image distributions persists. While domain generalization is acknowledged as vital for robust application in clinical settings, the challenges stemming from training with a limited Field of View (FOV) remain unaddressed. This limitation leads to false predictions when applied to body regions beyond the FOV of the training data. In response to this problem, we propose a novel loss function that penalizes predictions in implausible body regions, applicable in both single-dataset and multi-dataset training schemes. It is realized with a Body Part Regression model that generates axial slice positional scores. Through comprehensive evaluation using a test set featuring varying FOVs, our approach demonstrates remarkable improvements in generalization capabilities. It effectively mitigates false positive tumor predictions up to 85% and significantly enhances overall segmentation performance.

nnU-Net Revisited: A Call for Rigorous Validation in 3D Medical Image Segmentation

The release of nnU-Net marked a paradigm shift in 3D medical image segmentation, demonstrating that a properly configured U-Net architecture could still achieve state-of-the-art results. Despite this, the pursuit of novel architectures, and the respective claims of superior performance over the U-Net baseline, continued. In this study, we demonstrate that many of these recent claims fail to hold up when scrutinized for common validation shortcomings, such as the use of inadequate baselines, insufficient datasets, and neglected computational resources. By meticulously avoiding these pitfalls, we conduct a thorough and comprehensive benchmarking of current segmentation methods including CNN-based, Transformer-based, and Mamba-based approaches. In contrast to current beliefs, we find that the recipe for state-of-the-art performance is 1) employing CNN-based U-Net models, including ResNet and ConvNeXt variants, 2) using the nnU-Net framework, and 3) scaling models to modern hardware resources. These results indicate an ongoing innovation bias towards novel architectures in the field and underscore the need for more stringent validation standards in the quest for scientific progress.

Skeleton Recall Loss for Connectivity Conserving and Resource Efficient Segmentation of Thin Tubular Structures

Accurately segmenting thin tubular structures, such as vessels, nerves, roads or concrete cracks, is a crucial task in computer vision. Standard deep learning-based segmentation loss functions, such as Dice or Cross-Entropy, focus on volumetric overlap, often at the expense of preserving structural connectivity or topology. This can lead to segmentation errors that adversely affect downstream tasks, including flow calculation, navigation, and structural inspection. Although current topology-focused losses mark an improvement, they introduce significant computational and memory overheads. This is particularly relevant for 3D data, rendering these losses infeasible for larger volumes as well as increasingly important multi-class segmentation problems. To mitigate this, we propose a novel Skeleton Recall Loss, which effectively addresses these challenges by circumventing intensive GPU-based calculations with inexpensive CPU operations. It demonstrates overall superior performance to current state-of-the-art approaches on five public datasets for topology-preserving segmentation, while substantially reducing computational overheads by more than 90%. In doing so, we introduce the first multi-class capable loss function for thin structure segmentation, excelling in both efficiency and efficacy for topology-preservation.

Embarrassingly Simple Scribble Supervision for 3D Medical Segmentation

Traditionally, segmentation algorithms require dense annotations for training, demanding significant annotation efforts, particularly within the 3D medical imaging field. Scribble-supervised learning emerges as a possible solution to this challenge, promising a reduction in annotation efforts when creating large-scale datasets. Recently, a plethora of methods for optimized learning from scribbles have been proposed, but have so far failed to position scribble annotation as a beneficial alternative. We relate this shortcoming to two major issues: 1) the complex nature of many methods which deeply ties them to the underlying segmentation model, thus preventing a migration to more powerful state-of-the-art models as the field progresses and 2) the lack of a systematic evaluation to validate consistent performance across the broader medical domain, resulting in a lack of trust when applying these methods to new segmentation problems. To address these issues, we propose a comprehensive scribble supervision benchmark consisting of seven datasets covering a diverse set of anatomies and pathologies imaged with varying modalities. We furthermore propose the systematic use of partial losses, i.e. losses that are only computed on annotated voxels. Contrary to most existing methods, these losses can be seamlessly integrated into state-of-the-art segmentation methods, enabling them to learn from scribble annotations while preserving their original loss formulations. Our evaluation using nnU-Net reveals that while most existing methods suffer from a lack of generalization, the proposed approach consistently delivers state-of-the-art performance. Thanks to its simplicity, our approach presents an embarrassingly simple yet effective solution to the challenges of scribble supervision. Source code as well as our extensive scribble benchmarking suite will be made publicly available upon publication.