Abstract

Automated detection of knee pathologies from MRI requires integrating evidence across three orthogonal imaging planes while handling variable scan lengths and class imbalance. ACL tears, meniscus tears, and general structural abnormalities are among the most common orthopedic presentations worldwide, yet radiological interpretation remains time-intensive and subject to inter-reader variability. We present KneeDINO, a multi-label framework that simultaneously detects all three pathologies from axial, coronal, and sagittal volumes. Our central contribution is a per-task cross-plane attention fusion pipeline: three independent RadioDINO ViT-B/16 encoders produce task-specific slice embeddings via per-task attention-pooling heads, and a shared cross-plane Transformer encoder is applied once per pathology, allowing each task to attend across planes using only the slice evidence most relevant to that task. A learnable per-plane embedding distinguishes axial, coronal, and sagittal tokens, which position-free self-attention alone cannot. Attention analysis reveals that all three tasks attend most strongly to the coronal plane, consistent with its clinical importance for joint-space evaluation, with no task exhibiting diagonal-dominant self-attention — confirming genuine cross-plane information exchange. A two-stage explainability pipeline passes volumetric EigenCAM saliency overlays to Gemini, producing plain-language summaries grounded in task-specific visual evidence. KneeDINO achieves AUC of 0.942 (ACL), 0.871 (meniscus), and 0.955 (abnormal) on MRNet, surpassing prior approaches on meniscus and abnormality detection while remaining competitive on ACL classification.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/EMERGE_021.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: Not Submitted

Link to Open Review

Open Review Page: https://openreview.net/forum?id=udRDkyCR20

BibTex

@InProceedings{JosRag_KneeDINO_MICCAISAT2026,
        author = { Joshi, Raghav},
        title = { { KneeDINO: Cross-Plane Attention Fusion for Multi-Label Knee Pathology Detection with Grounded LLM Explainability } },
        booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026 Workshops and Challenges},
        year = {2026},
        publisher = {Springer Nature Switzerland},
        volume = {LNCS 17266},
        month = {pending},
        page = {pending}
}


back to top