Fully automated segmentation of subcortical structures in CT head scans using deep learning

Abstract

Within the subcortex of the human brain, there are many subcortical anatomies that hold key roles in many fundamental physiological functions. These anatomies exhibit volumetric and morphological changes during the development of neurodegenerative disorders. Thus, it would be useful to perform subcortical segmentation to study their potential as biomarkers for neurological conditions. However, their small sizes make manual segmentation challenging and time-consuming. While numerous automated subcortical segmentation studies have been conducted in MR modality, there is a lack of for CT modality. Given that CT is more accessible, affordable and faster, we aim to develop tools and frameworks for automated subcortical segmentation in CT modality.

To address the lack of publicly available CT subcortical segmentation datasets, cross-domain label propagation from MR to CT was performed through our proposed automated framework of ensembling established MR segmentation models. The proposed framework was validated against manually-annotated datasets to ensure its robustness. The usefulness of our generated CT subcortical segmentation dataset was also validated using transfer learning. Subsequently, SOTA deep learning models, including Swin UNETR and nnU-Net, were trained to perform CT subcortical segmentation and their performance was evaluated.

Our proposed novel automated subcortical label generation framework, ASLGF, achieved increased robustness compared to the MR models standalone. Using our proposed framework, we generated a CT subcortical segmentation dataset that will be the first publicly available CT subcortical segmentation dataset and will enable researchers to develop models by training on it. Our trained deep learning models also achieved commendable average dice scores of 0.741 to 0.901. The code and dataset is made available on the Biomedical Computing Group’s Github: https://github.com/SCSE-Biomedical-Computing-Group/FYP_Augustine

Through this paper, we provided key contributions along various points of the automated segmentation process. Through the various toolkits and insights developed from this study, we hope to encourage and democratise more research in CT subcortical segmentation and in the broader scheme of things, to understand neurodegenerative disorders better.