aungmthein/oct-cartilage-cell-layer

OCT cartilage cell-layer (internal MVP)

Upload an OCT image of an engineered cartilage construct; the model returns a 3-class semantic segmentation overlay (above / cell layer / below) plus thickness statistics.

This is an internal-tool MVP, not a polished biologist-facing app. The production checkpoint is phase4-1f3f21ce_softmax_B__swin_umamba_fold4.pth trained on 68 labeled OCT images of engineered cartilage constructs.

Caveats

• —Out-of-distribution regime gap. Aggregate OOD Dice = 0.572 includes 15 no-bottom-regime images (cell layer extends below the visible frame) that the model was never trained for. The restricted-OOD metric on the thin-bounded subset is closer to ~0.81. See the model card for the regime breakdown.
• —Cold start ~15-20s on first request after Space sleeps; subsequent requests in the same GPU session run in ~6-11s.
• —CPU deploy is not supported: the encoder's selective-scan op is a CUDA-only kernel (mamba-ssm). ZeroGPU is mandatory.

Source

Model weights live in `aungmthein/oct-cell-layer-fold4-swin-umamba`.