Automatic Segmentation of Achilles Tendon Tissues Using Deep Convolutional Neural Network

Abstract

The automatic segmentation of Achilles tendon tissues is one of the preliminary steps towards creating a tool for diagnosing, prognosing, or monitoring changes in tendon organization over time. Manual delineation is the current approach of identifying Achilles region-of-interest (ROI), it is a tedious and time-consuming task. In this respect, the current work describes the first steps taken towards creating an automatic approach for Achilles tendon segmentation that utilize the capabilities of Deep Convolutional Neural Networks (CNNs). Firstly, the dataset has been pre-processed and manually segmented to be used as the ground-truth in the training and testing of the proposed automated model. Secondly, the model was trained and validated using three CNN architectures SegNet, ResNet-18 and ResNet-50. Finally, Tversky loss function, 3D augmentation and network ensembling approaches were used to improve the segmentation performance and to tackle challenges such as the limited size of the training dataset and data imbalance. The proposed fully automated segmentation method reached average Dice score of 0.904. In conclusion, this novel study demonstrates that a CNN approach is useful for performing accurate Achilles tendon segmentation in musculoskeletal imaging.