---
license: other
library_name: onnxruntime
pipeline_tag: image-segmentation
tags:
  - cardiac-mri
  - lge
  - myocardium-segmentation
  - scar-segmentation
  - onnx
---

# CMR Annotator JDL Myocardium and Scar Segmentation

Weight-only repository for the browser-side Joint Deep Learning (JDL)
myocardium and scar segmentation modules in CMR Annotator.

The two-stage scar pipeline:

1. Segment myocardium, either with `jdl_myocardium.onnx` or another compatible
   CMR Annotator myocardium provider.
2. Resize and soften the myocardium mask, multiply it with the normalized LGE
   image, and pass the weighted image to `jdl_scar.onnx`.
3. Decode the scar logits and map the result back into the source image space.

Both models accept float32 tensors shaped `[N, 1, 224, 224]` under the input
name `image` and return float32 logits shaped `[N, 2, 224, 224]` under the
output name `logits`.

## Checkpoints

| File | Size | SHA-256 |
| --- | ---: | --- |
| `checkpoints/jdl_myocardium.onnx` | 421.7 MB | `acaa352d106268c581b84da3a2627a96ee755d6488a5d6b0421376ca4997d17d` |
| `checkpoints/jdl_scar.onnx` | 421.7 MB | `0589305742e065b3c7410b6934a6e93955825dc77ebefe30da915fe8e5703ee0` |

These checkpoints are large and are downloaded only when their corresponding
CMR Annotator action or dependency is first run. CMR Annotator verifies the
declared SHA-256 checksum and caches the verified bytes in IndexedDB.

## Intended use

These research checkpoints produce candidate annotations for expert review;
they are not a standalone clinical diagnostic device. The original JDL
myocardium model is short-axis biased. CMR Annotator therefore uses its newer
all-view myocardium provider by default and exposes JDL myocardium as an
explicit alternative.