DisMix: Order-Aware Mixup for Medical Imaging via Disentangling Ordinal and Non-Ordinal Features

2026-08-05

AI Summary

Standard image mixup conflates disease severity (an ordinal attribute) with appearance variations (non-ordinal attributes) in medical ordinal classification tasks, thereby disrupting the inherent label order structure. This work proposes DisMix, the first mixup-based augmentation framework that explicitly disentangles ordinal and non-ordinal features. DisMix employs a dual-codebook vector-quantized variational autoencoder to separate these two feature types and applies distinct augmentation strategies—interpolation for ordinal features and random replacement for non-ordinal ones—to generate samples that preserve clinical grading order while enhancing appearance diversity. Evaluated across four medical imaging datasets, DisMix consistently outperforms six mixup baselines combined with six ordinal classifiers, demonstrating robust performance under data scarcity and variations in clinical grading protocols.

Abstract

Image mixup is a widely adopted data augmentation strategy, yet it is ill-suited for ordinal classification tasks such as medical disease grading, where labels encode a progression of severity. By indiscriminately blending disease-severity cues (ordinal) with appearance-level variation (non-ordinal), standard mixup produces samples that distort the very ordinal structure that underpins clinical severity grading. We introduce DisMix, an order-aware mixup framework for ordinal classification. DisMix disentangles ordinal and non-ordinal features via a dual-codebook VQ-VAE, allowing each subspace to be mixed independently: ordinal codes are interpolated to produce meaningful intermediate ranks, while non-ordinal codes are varied to introduce appearance diversity without corrupting the ordinal signal. Across four medical imaging datasets, DisMix shows the best aggregate performance among six image mixup baselines paired with six ordinal classifiers and remains effective under data scarcity and clinical grading variability.

Problem

• ordinal classification
• medical imaging
• data augmentation
• mixup
• disease grading

Innovation

• order-aware mixup
• disentangled representation
• ordinal classification
• medical imaging
• VQ-VAE