custom_ops XRV DenseNet121-res224-all

Multi-label chest X-ray classifier (18 pathologies, TorchXRayVision densenet121-res224-all). Two paths: (1) Drop a CXR image for live in-browser preprocess + inference (xrv normalize + center crop + 224 resize, all in the worker via shared decodeImageToFloat32); or (2) Run verify + bench against precomputed PyTorch references for parity validation.

Classify your CXR

Drop a chest X-ray PNG / JPG here, or .

Pathology scores (top-5)

Verification harness

Runs precomputed inputs x.bin (synthetic seed=0) and x_cxr.bin (real test_data/test_cxr.png after server-side xrv preprocess) through the WebGPU forward, applies sigmoid + op_norm, and diff-checks against PyTorch references.
idle

Real CXR — ours (top-5 post sigmoid+op_norm)

Real CXR — PyTorch reference

Synthetic seed=0 — ours

Synthetic seed=0 — PyTorch

Log