Automated neural network femur segmentation performance in computed tomography images of older adults with obesity

Femur segmentation is a precursor to image analysis pipelines that evaluate hip bone measures with subject-specific finite element models, but historically required time-intensive efforts of operators. Implementing deep learning techniques offers a promising pathway for rapid, automated, and accurate segmentation. This study evaluates the feasibility of using a convolutional neural network (CNN) developed and validated for femur segmentation of Icelandic older adults to segment femurs in a sample of 166 CT scans of older adults with obesity in the United States. The performance of the segmentation model was quantitatively evaluated against manually segmented ground truth data using the Dice similarity coefficient (DSC) and the 95% Hausdorff distance (HD95), and qualitatively evaluated using manual image review. The mean DSC was 0.974 ± 0.009 and mean HD95 for the foreground was 1.078 ± 0.41 mm, indicating excellent segmentation quality. The CNN segmentation model averaged 32 ± 3 s for each mask prediction. Visual inspection revealed segmentations with minor errors in delineating boundaries at the femoral head and detecting osteophytes, requiring refinement during post-processing. The present CNN, when evaluated on this dataset of older adults with obesity, produced femur segmentations with improved quality and speed compared to manual segmentation. Additional model training with a more diverse training dataset could help further minimize manual intervention that is currently required during pre- and postprocessing. This study demonstrates the potential for femur segmentation CNN models to be applicable to diverse clinical datasets.