calmar: SEALS fold-0 fp16 ONNX DWI/ADC lesion model

calmar/models/README.md

@@ -0,0 +1,21 @@
+# CALMaR lesion models
+
+## lnm-deepisles-seals-fold0-128-fp16.onnx
+
+DWI/TRACE + ADC acute ischaemic stroke lesion segmentation for CALMaR's
+editable lesion seed.
+
+- Source: SEALS, the nnU-Net v1 member of DeepISLES (ISLES'22 winner),
+  fold 0 (`model_final_checkpoint.model`) from the DeepISLES weights,
+  Zenodo record 14026715 (`stroke_ensemble_weights.7z`,
+  md5 be5b6dfcd66b55c2e6dc6db9a5880f7f), Apache-2.0.
+- Paper: de la Rosa et al., "DeepISLES: a clinically validated ischemic
+  stroke segmentation model from the ISLES'22 challenge", Nature
+  Communications (2025).
+- Conversion: `apps/calmar/scripts/convert_seals_model.py` in
+  github.com/neurodesk/webapps. Static 1x2x128x128x128 input
+  (channel 0 = DWI/TRACE, channel 1 = ADC, 1 mm), output logits
+  (background, lesion), opset 17, fp16 weights with fp32 inputs/outputs.
+- Inference: `apps/calmar/scripts/run_seals_onnx.py` (nnU-Net v1
+  preprocessing, step-0.5 Gaussian sliding window, 8 mirrorings).
+- sha256 54c01314ea7ad26b092266ac6cb39bc9f35ac84d8983186af734f05b1cd1c329

calmar/models/lnm-deepisles-seals-fold0-128-fp16.onnx

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:54c01314ea7ad26b092266ac6cb39bc9f35ac84d8983186af734f05b1cd1c329
+size 62429252