mamounyosef/knee-abnormality-detection-dinov3

Knee Abnormality Detection: DINOv3 models (RSNA, Kaggle)

Built with DINOv3. The two DINOv3 ConvNeXt-Base fine-tunes from my solution to the RSNA Knee Abnormality Detection Kaggle competition. They are kept apart from the main model repo because fine-tuned DINOv3 weights must be distributed under the DINOv3 License (copy in LICENSE.md).

Model	What	Labels	Val	Audit-58	Alone: public / private
dinov3_convnext_base_v5	DINOv3 ConvNeXt-Base, per-slice at 512 px, on v4 labels	v4	0.9211	0.926	-
dinov3_convnext_base	DINOv3 ConvNeXt-Base, per-slice at 512 px, on v2 labels	v2	0.9011	0.8957	-

• dinov3_convnext_base is a member of the 14-model ensemble (public 0.942 / private 0.928).
• dinov3_convnext_base_v5 is a member of the 4-model ensemble (public 0.942 / private 0.927).

Each folder has the weights, config, a standalone modeling.py, metrics, the training log and a full model card. Code and write-up: GitHub.

Licence and intended use

• Weights: DINOv3 License (LICENSE.md). Acknowledge DINOv3 in any publication that uses them.
• Trained on RSNA competition data, licensed for research use only (RSNA MIRA terms). No competition images, labels or reports are included.
• Research use only. Not a medical device and not for clinical use.