FindXRay encoder: RAD-DINO (224 px), ONNX int8

The image encoder of FindXRay, which finds similar past chest X-rays. It runs in the visitor's browser with ONNX Runtime Web.

Derived from microsoft/rad-dino (MIT licence): exported to ONNX and compressed with ONNX Runtime dynamic int8 quantisation (330 MB -> 86 MB).
Input pixel_values [1, 3, 224, 224]: grayscale X-ray, short side resized to 518 then 224 (bicubic), centre-cropped to 224, scaled to [0, 1], normalised with mean 0.5307 and std 0.2583, repeated over 3 channels.
Output embedding [1, 768]: L2-normalised CLS embedding.
Quality: cosine 0.996 vs the full-precision model; retrieval on IU X-ray unchanged (P@5 0.417).

Research demo only. Not a medical device and not for diagnosis.

Downloads last month: -; Downloads are not tracked for this model. How to track

Inference Providers NEW

This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Model tree for tush-2ar/findxray-encoder

Base model

Quantized

(2)

this model