Automatic Segmentation Of Bone And Muscle Structures In Ct Volumes Using Convex Relaxation And Fine-Tuning

Segmentation of muscle and bone structures in CT volumes is a complicated task for physicians and surgeons due to the similarity of Hounsfield values with those within surrounding organs. The lack of efficient and automatic tools for this task makes that most physicians and surgeons prefer to segment bone and/or muscle tissue manually rather than working with specific-focused-tissue segmentation tools. In this work, an automatic tool for the joint segmentation of muscle and bone structures is described. The tool implements the segmentation in two stages. In the first stage an energy minimization approach via convex relaxation is performed. In the second step, a more accurate muscle segmentation is achieved using probability maps. A public database of CT volumes have been used to test the algorithm and different metrics, including Dice, Jaccard, sensitivity and specificity, have been calculated to evaluate the algorithm performance. The results provided outperform recent methods using the same public database.