us-ijepa-v3-encoder

I-JEPA ViT-B/16 encoder for breast ultrasound, from an MSc thesis. Encoder weights only (encoder state dict and the training cfg), 343 MB, float32. Research use; not a clinical tool.

• Input: 224 x 224 grayscale ultrasound, repeated to 3 channels, scaled to [0, 1].
• Output: patch tokens [B, 196, 768]; the thesis probes use the mean over patches.
• Frozen linear probe, BUS-BRA (patient-grouped split): validation AUROC 0.814, test 0.870 on whole frames.

Used by the agent repository, which loads it from checkpoints/us_ijepa_v3_encoder.pt.