Method of muscle tissue segmentation in computed tomography images based on preprocessed three-channel images

Abstract

The results of a study of a preprocessing influence method based on the formation of three-channel images on the accuracy of muscle tissue segmentation models on the computed tomography scans corresponding to the levels of the vertebrae of the thoracic and lumbar spine are presented. Ten models have been trained and tested on the Sparsely Annotated Region and Organ Segmentation dataset. The values of the Dice similarity coefficient and the Intersection over Union in the ranges of 0.9353–0.9421 and 0.8737–0.8885 were obtained. The use of a three-channel approach to the formation of input data increased the accuracy of models of four of the five architectures considered. Trained models can be used to quickly and accurately annotate muscle tissue during the diagnostic process.

Keywords

• computer vision
• segmentation
• computed tomography
• muscle tissue
• diagnostics
• u-net