DisMix: Order-Aware Mixup for Medical Imaging via Disentangling Ordinal and Non-Ordinal Features
A novel data augmentation method named DisMix has been developed to overcome the shortcomings of conventional mixup strategies in ordinal classification, especially in the context of medical disease grading. This technique, outlined in an arXiv paper (2608.04652), separates ordinal from non-ordinal features by utilizing a dual-codebook VQ-VAE, which enables the independent mixing of each feature subspace. It interpolates ordinal codes to create intermediate severity levels, while non-ordinal codes are adjusted to enhance appearance diversity without compromising the ordinal information. Tested on four medical imaging datasets, DisMix outperformed six image mixup baselines when used with six ordinal classifiers, marking a significant advancement for accurate severity assessment in medical imaging essential for effective diagnosis and treatment planning.
Key facts
- DisMix is an order-aware mixup framework for ordinal classification.
- It uses a dual-codebook VQ-VAE to disentangle ordinal and non-ordinal features.
- Ordinal codes are interpolated to produce meaningful intermediate ranks.
- Non-ordinal codes are varied to introduce appearance diversity without corrupting the ordinal signal.
- Evaluated on four medical imaging datasets.
- Shows best aggregate performance among six image mixup baselines paired with six ordinal classifiers.
- Paper available on arXiv with ID 2608.04652.
- Standard mixup distorts ordinal structure in medical disease grading.
Entities
Institutions
- arXiv