Medical image classification improved by separating disease and other factors
Decoupling Disease, Covariates, and Individual Variability: A Unified Disentanglement Framework for Medical Image Classification
Computer Vision and Pattern Recognition
Summary
Medical images can be hard to analyze because factors like age, gender, and individual differences mix with disease signals. The authors developed a new method called MedIDL that splits these factors into separate parts to better identify disease patterns. They tested this method on many types of medical images and found it works better than existing approaches. It also helps point out the important features doctors use to diagnose diseases.
What this means in practice
- •For medical imaging software developers: Develop diagnostic tools that more accurately identify diseases by separating disease features from patient differences.$Commercial implications: Enables commercial diagnostic imaging software to deliver more precise disease detection across diverse patient groups.
- •For clinical data analysts: Improve disease analysis by isolating disease-specific signals from confounding factors in medical imaging datasets.
Authors
Shengjie Zhang, Jinglin Zhang, Zhuangzhuang Jiang, Ziqi Yu, Yipin Zhang, Qi Zhang, Xiang Chen, Haibo Yang, Fei Gao, Longbiao Cui, Yuan Zhou, Xiao-Yong Zhang, Alzheimer's Disease Neuroimaging Initiative
Abstract
Accurately isolating disease-related features from confounding covariates (e.g., age, gender, site) and individual variations remains a fundamental challenge in medical image classification. Traditional regression-based approaches may ignore non-linear relations between image features and true covariates. To overcome this issue, we present a generalized Medical Imaging Disentanglement Learning (MedIDL) framework. MedIDL maps image features into three mutually orthogonal latent spaces through specialized disentanglement heads: a disease classification head guided by a supervised loss, a covariate-alignment head constrained by cross-subject similarity matching, and a Gaussian head absorbing individual variations. We evaluated our framework across 7 datasets encompassing diverse imaging modalities. MedIDL outperforms state-of-the-art supervised and self-supervised classification methods in accuracy across all datasets. Association analyses demonstrate that MedIDL successfully isolates target-specific latent representations. Gradient-based interpretability mappings localize pathognomonic patterns aligning with established clinical literature.