Can Score-Based Generative Modeling Effectively Handle Medical Image Classification?

TL;DR

Problem: medical image classification is challenged by limited data and overlapping class distributions, which can hinder discriminative approaches. Approach: the authors adopt a score-based generative diffusion framework conditioned on class labels to learn class-conditional data distributions and compute likelihoods via forward-backward dynamics. Contributions: first demonstration of score-based generative classifiers for medical images, with strong results on CBIS-DDSM, INbreast, and VinDr-Mammo, and a public code release. Significance: the method provides a principled likelihood-based alternative for medical image classification that can extend to segmentation and other tasks, potentially improving reliability under data scarcity and imbalance.

Abstract

The remarkable success of deep learning in recent years has prompted applications in medical image classification and diagnosis tasks. While classification models have demonstrated robustness in classifying simpler datasets like MNIST or natural images such as ImageNet, this resilience is not consistently observed in complex medical image datasets where data is more scarce and lacks diversity. Moreover, previous findings on natural image datasets have indicated a potential trade-off between data likelihood and classification accuracy. In this study, we explore the use of score-based generative models as classifiers for medical images, specifically mammographic images. Our findings suggest that our proposed generative classifier model not only achieves superior classification results on CBIS-DDSM, INbreast and Vin-Dr Mammo datasets, but also introduces a novel approach to image classification in a broader context. Our code is publicly available at https://github.com/sushmitasarker/sgc_for_medical_image_classification

Figures (2)

• Figure 1: Illustration of score-based approach for (binary) classification task. Class A and B represents two distinct classes in the data distribution space (left), while the score function through denoising score matching represents the direction towards high density regions of respective class (right).
• Figure : Training Algorithm