N/A

The goal is to develop and validate a 2.5D deep learning neural network (DLNN) to automatically classify thigh muscle into 10 classes and evaluate its classification accuracy over 2D DLNN.

The clinical dataset consists of 48 thigh volume(TV) cropped from 24 anonymized non-contrast CT DICOM of lower extremities. Cropped volumes were aligned with femur axis and resample in 2mm voxel spacing. To reduce the annotation workload, final expert ground truth annotation was created by editing the predicted labels of muscle by a newly developed stacked U-Net DLNN. Stacked U-Net produces relatively higher segmentation accuracy on smaller muscles even when it is trained with small number annotated datasets. Proposed 2.5D DLNN consists of three 2D U-Net(optimizer: Adam, lr=1e-4,decay=1e-3) trained with axial, coronal and sagittal muscle slices respectively. A voting algorithm was used to combine the output of 2D U-Nets to create final segmentation. 2.5D U-Net was trained on PC(Intel Xeon 2.20GHz 128GB, NVIDIA Tesla P100-SXM2-16GB) with 38 TV(Epoch:100, Batch:32) and the remaining 10 TV were used to evaluate segmentation accuracy of 10 classes within Thigh. The result segmentation of both left and right thigh were de-cropped to original CT volume space. Finally, segmentation accuracies were compared between proposed DLNN and 2D U-Net(axial).

Average segmentation DSC score accuracy of all classes with 2.5D U-Net as 91.18% and Housdorff distance(HD) as 17mm. We found DSC score for 2D U-Net was 2.9% lower and HD was more than four times higher than the that of 2.5D U-Net.

Successfully implemented end-to-end solution for complete automatic classification with reasonable accuracy of thigh muscle into 10 classes . The same could be easily extend to muscle segmentation of any other body parts (lower limb, arm, shoulder etc. ). To date, there is no other study of deep machine learning algorithm used except our study for CT based semantic muscle segmentation.

Muscle segmentation functionality on PACS may improve visibility and can enable automatic quantitative evaluation of muscle atrophy with the disease progression. Change in volume or shapes of muscles will enable therapeutic interventions to be targeted to the affected regions only.