COJEPA: New AI framework achieves 0.84 twin recall on brain MRI

COJEPA (Contrastive Joint-Embedding Prediction) extends I-JEPA to 3D brain MRI by introducing foreground-aware block masking, hierarchical convolutional patch embeddings, and world-space sinusoidal positional encodings. The model jointly optimizes a local predictive loss (JEPA) and a global contrastive loss, enabling it to learn both fine-grained anatomical structure and whole-brain discriminability. Training on two large cohorts—HCP-YA and AABC (total N=2,286, ages 22-90)—demonstrates the framework's ability to capture meaningful representations without any labeled data.

The results highlight COJEPA's versatility: it achieves the best monozygotic twin retrieval accuracy (0.84 at rank@1), indicating strong preservation of genetic similarity patterns. For age regression on the OpenBHB 3.0T dataset, it sets a new state-of-the-art mean absolute error of 2.55 years. On BraTS 2024 tumor segmentation, COJEPA matches the performance of purely contrastive methods on whole-tumor Dice, while outperforming them on the combined objective. This work shows that combining predictive and contrastive objectives produces representations that are simultaneously locally structured and globally discriminative—critical for downstream medical imaging tasks where labeled data is scarce.