Deep Learning for Segmentation and Landmarking of the Human Anatomy: Applying nnU-Net and SpatialConfiguration-Net to Automate Statistical Shape Modeling for Sternum and Costal Cartilage

Abstract

During vehicle development, simulations are used in part to replicate the mechanical response of the human body during crash testing. To ensure realism, human body models are employed. However, current models often lack an accurate representation of thoracic anatomy, including the sternum and costal cartilage. One approach to improve these models is to develop a statistical shape model based on computed tomography (CT) scans of the thoracic region. The creation of statistical shape models from CT data involves multiple steps and is typically time-consuming. This study focuses on automating the first two steps of the process: segmentation and landmarking, using two different deep learning models. The aim is to enhance efficiency and consistency compared to manual processing. Additionally, the impact of dataset size on the models’ performance is investigated. The creation of statistical shape models from CT data involves multiple steps and is typically time-consuming. This study focuses on automating the first two steps of the process: segmentation and landmarking using deep learning. The aim is to enhance efficiency and consistency compared to manual processing. Additionally, the impact of dataset size on model performance is investigated. The methodology includes preparing datasets with manual segmentations and landmarks, formatting them for use with deep learning models, and training two architectures: nnU-Net for segmentation and SpatialConfiguration-Net (SCN) for landmarking. The segmentation results from nnU-Net are generally accurate and more consistent than manual segmentations, though some manual post-processing is still necessary. SCN also demonstrates promising performance for landmarking, with similar requirements for manual correction. Both models demonstrate improved performance as the size of the training dataset increases. Additionally, nnU-Net’s segmentation accuracy improves when trained on datasets with more precise annotations. Despite the small dataset sizes (up to 10 scans for segmentation and 20 for landmarking), both models achieved promising results, highlighting their robustness even with limited data. In conclusion, deep learning can effectively automate the segmentation and landmarking processes required for statistical shape model development, offering improvements in both efficiency and consistency. The findings also suggest that model performance continues to improve with larger and more diverse datasets.