Machine learning models for medical image analysis often suffer from poor performance on important subsets of a population that are not identified during training or testing. For example, overall performance of a cancer detection model may be high, but the model still consistently misses a rare but aggressive cancer subtype. We refer to this problem as hidden stratification, and observe that it results from incompletely describing the meaningful variation in a dataset. While hidden stratification can substantially reduce the clinical efficacy of machine learning models, its effects remain difficult to measure. In this work, we assess the utility of several possible techniques for measuring and describing hidden stratification effects, and characterize these effects on multiple medical imaging datasets. We find evidence that hidden stratification can occur in unidentified imaging subsets with low prevalence, low label quality, subtle distinguishing features, or spurious correlates, and that it can result in relative performance differences of over 20% on clinically important subsets. Finally, we explore the clinical implications of our findings, and suggest that evaluation of hidden stratification should be a critical component of any machine learning deployment in medical imaging.
DNA Methylation Profiling in Melanoma: From Lesion Classification to Therapeutic Stratification
German Cancer Research Center (DKFZ)（德国癌症研究中心）；University Heidelberg（海德堡大学）；Technische Universität Dresden（德累斯顿工业大学）；University Hospital (UKSH), Kiel（基尔大学医院）；LMU Munich（慕尼黑大学）；Uniklinikum Erlangen（埃尔朗根大学医院）；Bavarian Cancer Research Center (BZKF)（巴伐利亚癌症研究中心）；Charité – Universitätsmedizin Berlin（柏林夏里特医学院）；Freie Universität Berlin（柏林自由大学）；Humboldt-Universität zu Berlin（柏林洪堡大学）；Vivantes Hospital Spandau（施潘道维万特斯医院）；Dermpath München（慕尼黑皮肤病理实验室）；University Hospital Regensburg（雷根斯堡大学医院）；University Hospital Essen（埃森大学医院）；Deutsches Zentrum Immuntherapie (DZI)（德国免疫治疗中心）
Deep Hypothesis Tests Detect Clinically Relevant Subgroup Shifts in Medical Images
Hertie Institute for Artificial Intelligence in Brain Health（赫蒂脑健康人工智能研究所）；Institute of Ophthalmic Research（眼科研究所）；Tübingen AI Center（图宾根人工智能中心）；University of Tübingen（图宾根大学）；Department of Pathology and Neuropathology（病理与神经病理科）；University Hospital Tübingen（图宾根大学医院）；Comprehensive Cancer Center Tübingen（图宾根综合癌症中心）；Cluster of Excellence Machine Learning: New Perspectives for Science（机器学习卓越集群：科学新视角）；Gatsby Computational Neuroscience Unit（盖茨比计算神经科学研究组）；University College London（伦敦大学学院）
On the notion of missingness for path attribution explainability methods in medical settings: Guiding the selection of medically meaningful baselines
Research Group MITI, Technical University of Munich（慕尼黑技术大学MITI研究组）；Chair for AI in Healthcare and Medicine, Technical University of Munich (TUM) and TUM University Hospital（慕尼黑技术大学人工智能在医疗健康中的chair及慕尼黑技术大学医院）；Department of Computing, Imperial College London（伦敦帝国理工学院计算机系）；Department of Surgery, Technical University of Munich（慕尼黑技术大学外科系）
Graph Representation Learning of Longitudinal Medical Imaging Trajectories for Treatment Response Prediction
Technical University of Munich（慕尼黑工业大学）；TUM University Hospital Rechts der Isar（慕尼黑工业大学伊萨尔河右岸大学医院）；Helmholtz Munich（慕尼黑亥姆霍兹中心）；Universitat de Barcelona（巴塞罗那大学）；Medical University of Vienna（维也纳医科大学）；King’s College London（伦敦国王学院）；Munich Center for Machine Learning（慕尼黑机器学习中心）
