A novel deep learning framework for conditioned feature modulation and synthetic image generation for severity-aware disease classification

Abstract

Deep learning systems have achieved substantial progress in visual recognition and image synthesis, but two structural limitations continue to restrict their use in fine-grained, data-scarce settings. First, multitask recognition models often predict related outputs independently from a shared representation, even when one prediction should condition the interpretation of another. Second, generative models used for data augmentation often rely on class labels or generic text prompts, which provide limited information about the fine-grained visual structure of the target category. These limitations are especially important when labelled data are imbalanced and when the target labels have an ordered severity structure. This thesis addresses these limitations through two complementary frameworks. The first, Disease-Conditioned Severity Calibration (DCSC), is a multitask recognition framework that conditions a secondary prediction pathway on a co-predicted primary label. DCSC uses a VMamba backbone, FiLM-based feature modulation, disease-conditioned severity prototypes, CORAL ordinal regression, and an Earth Mover’s Distance smoothing loss to make severity prediction context-aware rather than independent of the primary class prediction.