MADJEPA Architecture for Medical Anomaly Detection
November 20, 2026
MADJEPA uses a multi-stage joint-embedding predictive architecture (JEPA) to detect medical anomalies using only healthy patient data. The pipeline combines image-based JEPA training with target encoder distillation to prevent over-generalization to abnormal data.
HOW THIS AFFECTS YOU
●
researcherYou can build robust anomaly detectors using self-supervised JEPA training when labeled abnormal data is unavailable.
●
healthThis approach helps identify medical abnormalities using models trained exclusively on healthy patient images.