Current surface and fiber models perform poorly in compressed or highly curved regions. Improvements to these areas will yield immediate and significant improvements to current autosegmentation methods.
Current models are mostly trained using nnUNetv2 residual encoder UNets , with medial surface loss , dice loss, and binary cross-entropy. Most are trained using the nnUNetTrainerMedialSurfaceRecall located here which uses a modified version of MIC-DKFZ's skeleton recall loss
Surface models have been trained using this dataset which contains 1754 volume/label pairs of segmentation derived surface labels.
Fiber models have been trained primarily using this series of datasets, which continues to expand as our fiber annotation team continues working.
Input Format: image/label pairs of surface or fiber labels
Output Format: zarr
Result: High-accuracy predictions of written surfaces or fibers. Good fiber predictions would consist of long connected components without branching
Examples from Scroll 1:
Current surface and fiber models perform poorly in compressed or highly curved regions. Improvements to these areas will yield immediate and significant improvements to current autosegmentation methods.
Current models are mostly trained using nnUNetv2 residual encoder UNets , with medial surface loss , dice loss, and binary cross-entropy. Most are trained using the nnUNetTrainerMedialSurfaceRecall located here which uses a modified version of MIC-DKFZ's skeleton recall loss
Surface models have been trained using this dataset which contains 1754 volume/label pairs of segmentation derived surface labels.
Fiber models have been trained primarily using this series of datasets, which continues to expand as our fiber annotation team continues working.
Input Format: image/label pairs of surface or fiber labels
Output Format: zarr
Result: High-accuracy predictions of written surfaces or fibers. Good fiber predictions would consist of long connected components without branching
Examples from Scroll 1: