Automatic Segmentation of the Pectoralis Major Muscle in Magnetic Resonance Imaging Using nnU-Net: Evaluation of Training Mask Modifications and Anatomical Thickness Measurements

Abstract

Computational human body models (HBMs) can improve vehicle safety research by accounting for anatomical variability that is difficult to capture using traditional crash testing methods. Statistical shape models (SSMs) provide a means of representing this variability using medical imaging data, but their development requires accurate and reproducible anatomical segmentations. This thesis investigated the feasibility of automated segmentation of the pectoralis major muscle from magnetic resonance imaging (MRI) data using the nnU-Net framework. To the best of our knowledge, automated segmentation of the pectoralis major muscle from MRI data has not previously been investigated. The applicability of the segmentation methodology to other muscles was not evaluated and therefore remains to be investigated. Eleven manually segmented subjects from the TotalSegmentator MRI dataset were used to train and evaluate three training datasets: the original manually segmented masks (Nominal), masks systematically contracted by one voxel layer (Contracted), and masks systematically expanded by one voxel layer (Expanded). This comparison was performed to investigate how systematic modifications of the training masks affected pectoralis major segmentation performance. In addition, a repeatable landmark-based methodology was developed for measuring pectoralis major muscle thickness from the manually segmented MRI data, providing an initial anatomical baseline for future statistical shape modelling. The results indicate that the use of nnU-Net for automated segmentation of the pectoralis major muscle is feasible; however, due to the limited amount of available data, a highly accurate model could not be trained. The Expanded dataset produced slightly improved average segmentation performance compared with the Nominal dataset, whereas the Contracted dataset resulted in poorer performance. However, the differences between the Nominal and Expanded datasets were small relative to the variation between subjects, indicating that anatomical variability and limited dataset size had a greater influence on segmentation performance than the systematic mask modifications. The proposed pectoralis major segmentation and thickness measurement methodologies provide a foundation for future work on statistical shape models and human body models for vehicle safety applications.