Body Composition Assessment in 3D CT Images

Body composition as a diagnostic and prognostic biomarker is gaining importance in various medical fields such as oncology. Therefore, accurate quantification methods are necessary, such as analyzing CT images. While several studies introduced deep learning approaches to automatically segment a single slice, quantifying body composition in 3D remains understudied due to the high required annotation effort. This thesis explores the use of annotation-efficient strategies in developing a body composition segmentation model for the abdomen and pelvis. To address this, a fine-tuning strategy was proposed to optimize the annotation process and extend the model's generalization performance trained with L3 slices to the entire abdomen and pelvis. Moreover, a self-supervised pre-training using a contrastive loss was employed to leverage unlabeled data. The goal was to efficiently use the annotated data in developing the segmentation model. The results showed a significant acceleration of the annotation process. However, the pre-training added only limited benefits. The final model achieved excellent results with dice scores of SM: 0.96, VAT: 0.93, and SAT: 0.97. The insights gained from this study are useful to improve annotation procedures, and the developed body composition segmentation model can be used for further evaluation.