Multi-Task Learning Outperforms State-of-the-Art in Medical Anomaly Detection

Researchers propose MTL-MAD, a multi-task learning framework for anomaly detection in medical images. Unlike recent methods that rely on a single pretext task and large pre-trained models, MTL-MAD learns multiple self-supervised and pseudo-labeling tasks from scratch using a Mixture-of-Experts (MoE) joint model. By integrating diverse proxy tasks, the model captures robust representations of normal anatomy, enabling anomaly scoring based on task performance during inference. Experiments on the BMAD benchmark, covering multiple medical imaging modalities, show MTL-MAD surpasses all state-of-the-art competitors. The approach demonstrates that multi-task learners can be highly effective for medical anomaly detection without requiring pre-training.