Segmentation of hip joint anatomy structures from radiographic images

Abstract

This paper deals with the problem of a hip joint segmentation in radiographic images with the use of a deep learning approach. The paper is focused on training nnU-Net models and creating an original dataset that contains 150 radiographs, 100 training and 50 test images. There are six trained models, five from cross-validation training and one trained on all training data. All models are evaluated on the test dataset using the Dice score for individual labels and the combined mean Dice score for the image. The best-performing model was the model trained on all training images. The most challenging labels for segmentation were those representing the Köhler teardrop and the space between the femoral head, teardrop and acetabulum due to their size and variability observed across the dataset.