haruspex

A knee MRI model for the RSNA Knee Abnormality Detection challenge, flagging 12 findings from ligament and meniscus tears to effusion, contusion and fracture.

Category

Machine Learning

Industry

Medical Imaging

Role

Solo Competitor

Stack

Python, PyTorch, EfficientNet

Three knee MRI slices with highlighted regions and labels for findings such as ACL and meniscus
What I built

The model reads 2.5D windows of three slices from the sagittal and coronal planes, plus axial where it helps. A learned plane embedding tells it which view it is looking at, and attention pooling for each finding picks out the slices that matter for that label.

Challenge and solution

One network rarely wins on every finding. I trained EfficientNet B0 at 224 pixels and B2 at 256, with and without axial input, used mirrored coronal and axial augmentation, and combined them in an equal-weight rank-average ensemble.

Approach

Final scores of 0.919 on the public leaderboard and 0.909 on the private leaderboard. Findings covered include the ACL, MCL and menisci, osteoarthritis, effusion, synovitis, Baker's cyst, bone contusion and fracture, each scored separately.