Efficient cascaded V‐net optimization for lower extremity CTsegmentation validated using bone morphology assessment

Abstract

AbstractSemantic segmentation of bone from lower extremity computerized tomography(CT) scans can improve and accelerate the visualization, diagnosis, and surgicalplanning in orthopaedics. However, the large field of view of these scans makesautomatic segmentation using deep learning based methods challenging, slow andgraphical processing unit (GPU) memory intensive. We investigated methods tomore efficiently represent anatomical context for accurate and fast segmentationand compared these with state‐of‐the‐art methodology. Six lower extremity bonesfrom patients of two different datasets were manually segmented from CT scans,and used to train and optimize a cascaded deep learning approach. We varied thenumber of resolution levels, receptive fields, patch sizes, and number of V‐netblocks. The best performing network used a multi‐stage, cascaded V‐net approachwith 1283−643−323voxel patches as input. The average Dice coefficient over allbones was 0.98 ± 0.01, the mean surface distance was 0.26 ± 0.12 mm and the 95thpercentile Hausdorff distance 0.65 ± 0.28 mm. This was a significant improvementover the results of the state‐of‐the‐art nnU‐net, with only approximately 1/12th oftraining time, 1/3th of inference time and 1/4th of GPU memory required.Comparison of the morphometric measurements performed on automatic andmanual segmentations showed good correlation (Intraclass Correlation Coefficient[ICC] >0.8) for the alpha angle and excellent correlation (ICC >0.95) for the hip‐knee‐ankle angle, femoral inclination, femoral version, acetabular version, Lateral Centre‐Edge angle, acetabular coverage. The segmentations were generally of sufficientquality for the tested clinical applications and were performed accurately and quicklycompared to state‐of‐the‐art methodology from the literature