Preview

PDF, English - main document Download (21MB) | Lizenz: Lifelong Machine Learning for Biomedical Image Classification by Godau, Patrick underlies the terms of Creative Commons Attribution 4.0

Abstract

Despite rapid advances in the capabilities of AI, its application in healthcare faces unique challenges: stringent requirements to prove patient benefit, severe data scarcity, and shifting distributions across clinical environments.

These barriers span the entire AI lifecycle, requiring solutions at each stage. To address this, we propose a holistic Lifelong Learning framework that systematically addresses these challenges through three independent metacognitive loops - continuous improvements of the learning process itself: one to align validation efforts with clinical needs during the Design phase, another for effective knowledge transfer between tasks in the Develop phase, and a third to adapt models to changing environments in the Deploy phase.

Adding these loops to a learning system helps to overcome the challenges outlined above. First, we develop a structured interview process that captures the problem fingerprint of biomedical applications and enables automatic determination of appropriate performance measures aligned with clinical objectives. Second, we establish a method for quantifying task similarity and facilitating cross-institutional knowledge transfer while preserving patient data privacy. Our proposed bKLD measure underwent extensive evaluation across heterogeneous biomedical imaging tasks, setting new standards for task transferability estimation. Third, we comprehensively analyze prevalence shifts in deployment environments and propose a novel five-step workflow for model adaptation using only unlabeled samples from the deployment environment. Therein we quantify the present class prevalences and post-hoc re-calibrate a model, carefully considering the impact on decision rules and performance measures.

The experiments conducted demonstrate significant advancements in each area. Our recommendation process for aligning model performance metrics with actual clinical utility, reflects the consensus of an international consortium of 73 experts. Our knowledge transfer methodology allows the system to leverage experience from related tasks, exceeding previously proposed estimates of knowledge transferability in the most comprehensive benchmark we are aware of. Our prevalence shift compensation workflow prevents performance degradation across diverse biomedical imaging scenarios, enabling the system to automatically detect and adapt to changing environmental conditions without requiring new annotations.

This work represents the first comprehensive investigation of Lifelong Learning for biomedical image analysis, with tens of thousands of models trained and evaluated. By systematically leveraging metacognitive loops, we lay the groundwork for truly autonomous Lifelong Learning systems in healthcare that can continuously evolve in changing healthcare contexts.