Segmentation of muscle tissue in computed tomography images at the level of the L3 vertebra

With the increasing routine workload on radiologists associated with the need to analyze large numbers of images, there is a need to automate part of the analysis process. Sarcopenia is a condition in which there is a loss of muscle mass. To diagnose sarcopenia, computed tomography is most often used, from the images of which the volume of muscle tissue can be assessed. The first stage of the analysis is its contouring, which is performed manually, takes a long time and is not always performed with sufficient quality affecting the accuracy of estimates and, as a result, the patient's treatment plan. The subject of the study is the use of computer vision approaches for accurate segmentation of muscle tissue from computed tomography images for the purpose of sarcometry. The purpose of the study is to develop an approach to solving the problem of segmentation of collected and annotated images. An approach is presented that includes the stages of image pre-processing, segmentation using neural networks of the U-Net family, and post-processing. In total, 63 different configurations of the approach are considered, which differ in terms of data supplied to the input models and model architectures. The influence of the proposed method of post-processing the resulting binary masks on the segmentation accuracy is also evaluated. The approach, which includes pre-processing with table masking and anisotropic diffusion filtering, segmentation with an Inception U-Net architecture model, and post-processing based on contour analysis, achieves a Dice similarity coefficient of 0.9379 and Intersection over Union of 0.8824. Nine other configurations, the experimental results for which are reflected in the article, also demonstrated high values of these metrics (in the ranges of 0.9356–0.9374 and 0.8794–0.8822, respectively). The approach proposed in the article based on preprocessed three-channel images allows us to achieve metrics of 0.9364 and 0.8802, respectively, using the lightweight U-Net segmentation model. In accordance with the described approach, a software module was implemented in Python. The results of the study confirm the feasibility of using computer vision to assess muscle tissue parameters. The developed module can be used to reduce the routine workload on radiologists.