Equitable Medical Imaging via Cohort-Specialized Knowledge Distillation

Medical imaging classifiers can achieve high overall accuracy while exhibiting substantial performance disparities across demographic groups. We propose DIAGNOSE, a fairness-aware multi-teacher knowledge distillation framework designed to improve equitable performance without requiring sensitive attributes during deployment. DIAGNOSE follows a specialize-during-training, unify-for-deployment strategy. An ImageNet-pretrained Vision Transformer backbone is first fine-tuned using Fair Identity Scaling (FIS), which adaptively weights training samples according to classification difficulty and group-level fairness. The learned representation is then shared across cohort-specific teacher heads, enabling each teacher to capture specialized diagnostic patterns for its corresponding demographic group. Finally, knowledge from these specialized teachers is distilled into a single student using a hybrid objective that combines temperature-scaled knowledge distillation with fairness-weighted supervised learning. Sensitive demographic attributes are required only during training, while inference relies exclusively on medical images. DIAGNOSE therefore integrates cohort specialization, fairness-aware learning, and knowledge distillation into a unified framework for developing more equitable and deployment-ready medical imaging models.