Femoral head segmentation based on improved fully convolutional neural network for ultrasound images

Developmental dysplasia of the hip is a medical term representing the hip joint instability that appears mainly in infants. The assessment metric of physician is based on the femoral head coverage rate, which needs to segment the femoral head area in 2D ultrasound images. In this paper, we propose an approach to automatically segment the femoral head. The proposed method consists of two parts, firstly, mean filtering, morphological processing and least squares operation are used to detect the ilium and acetabular bone baseline to coarsely obtain the region of interest of the femoral head, then followed by an improved fully convolutional neural network named FNet which integrates the convolution encoder-decoder architecture, pooling indices and residual connection operation for more accurate segmentation. FNet is trained in a cascaded way, which can help the network learn more features with a limited dataset and thus further improve the segmentation performance. Experimental results show that the proposed method achieved an average dice, recall and IoU value of 0.946, 0.937 and 0.897. Moreover, the features learned by convolutional layers are visualized to demonstrate that FNet can focus on significant features, which is helpful to restore the contour of the femoral head more precisely. In conclusion, the proposed method is capable of segmenting femoral head accurately and guiding the diagnosis of developmental dysplasia of the hip.