Oct Cell Layer Fold4 Swin Umamba

--- library_name: pytorch tags: - image-segmentation - medical-imaging - oct - cartilage license: mit pipeline_tag: image-segmentation --- # OCT cartilage cell-layer (fold 4, Swin-UMamba) Production checkpoint for the OCT_Fabiola cell-layer segmentation task. 3-class semantic segmentation (above / cell layer / below) on 2D OCT images of engineered cartilage constructs. ## Architecture - Encoder: Swin-UMamba VMamba-Tiny, ImageNet-pretrained ([Liu et al., 2024](https://arxiv.org/abs/2402.03302)) - Head: 3-class softmax + per-pixel CE - Trained: 100 epochs, 384x384 random crops, fold 4 of 5-fold CV with construct-grouped splits ## Provenance - Run id: `phase4-1f3f21ce` - Trained 2026-05-21 on NVIDIA GB10 (DGX Spark) ## Numbers (5-fold CV + cross-batch OOD) | slice | n | Dice | MAE (px) | ECE | Brier | |---|---|---|---|---|---| | val (fold 4) | 14 | 0.84 | 32 | 0.05 | 0.13 | | OOD aggregate | 44 | 0.572 | 104.47 | 0.174 | 0.432 | | OOD restricted (thin-bounded subset) | 29 | ~0.81 | ~48 | -- | -- | | OOD no-bottom subset | 15 | ~0.10 | n/a | -- | -- | **Important caveat on the OOD numbers.** The OOD slice is bimodal-by-regime: 15 of 44 OOD images are in the no-bottom regime (cell layer extends below the visible frame), which is completely absent from training (0/68 training images). The model has zero training signal for that regime, so the aggregate OOD Dice 0.572 is dominated by an extrapolation failure, not a generalization failure on the in-distribution regime. The honest metric is the restricted-OOD number on the thin-bounded subset. ## Usage ```python from huggingface_hub import hf_hub_download ckpt_path = hf_hub_download( repo_id="aungmthein/oct-cell-layer-fold4-swin-umamba", filename="phase4-1f3f21ce_softmax_B__swin_umamba_fold4.pth", ) # Loader integration: see the demo Space at # https://huggingface.co/spaces/aungmthein/oct-cartilage-cell-layer ``` ## Inputs - 2D OCT image, single-channel grayscale recommended. - Cross-scale: the loader handles any H x W (zero-padded internally to a multiple of 32). - Recommended physical pixel spacing: 2.0 um/px (baseline 8000 x 1500 / 16 x 3 mm). ## Outputs - 3-class mask with values {0: above, 1: cell layer, 2: below}. - Thickness statistics (mean +/- std, percentiles, columns-measured count). ## License MIT (matches the upstream Swin-UMamba reference).

Source context: 0 downloads · 0 likes · Pipeline image-segmentation · Library pytorch · Repo aungmthein/oct-cell-layer-fold4-swin-umamba