Nouman Ahmad: Image Segmentation, Registration and Deep Regression: Methods for CT-based Large-Scale Body Composition Analysis

Abstract

Body composition refers to the amount and distribution of fat, muscle, bone, and other tissues in the body. It is associated with the risk of, for example, type 2 diabetes (T2D) and cardiovascular disease (CVD). Computed tomography (CT) is an imaging modality that provides detailed internal views, making it a valuable tool for body composition analysis.

CT imaging, however, exposes the imaged subjects to ionizing radiation, which poses health risks. To mitigate this, limited-slice protocols have been adopted for body composition studies. The Swedish CArdioPulmonary bioImage Study (SCAPIS) and Impaired Glucose Tolerance microbiota study (IGT) cohorts employ a three single-slice CT imaging approach. These cohort studies allow detailed investigations of the relationships between body composition, T2D, and CVD.

The primary aim of this thesis was to develop and evaluate different methods for detailed association studies of CT imaging and non-imaging data, applicable to large-scale body composition research. These methods have been employed to perform tasks such as image segmentation, image registration, voxel-wise association studies, deep regression, and classification.

An automatic segmentation method for the quantification of anatomical structures such as the liver, spleen, skeletal muscle, bone marrow, and various adipose tissue depots based on fully convolutional networks was developed and evaluated. To enable voxel-wise association studies, image registration was performed to align images from different subjects to a common reference space.

Additionally, deep learning-based convolutional neural networks were trained for image-driven regression tasks, predicting proof-of-concept measurements and more complex target variables properties such as age, BMI, and sex. In deep regression tasks, interpretability is important both to validate the model and to aid in the discovery of new associations between the image data and the target variable. To enable interpretability and trustworthiness, individual Gradient-weighted Class Activation Mapping (Grad-CAM) saliency maps were generated, deformed into a common reference space, and subsequently aggregated to produce cohort-level saliency maps. These maps highlight anatomical structures commonly associated with the predicted target (parameter) across the population. Furthermore, uncertainty quantification was employed to provide individual confidence intervals for each prediction, thereby estimating the quality of the model predictions.

In summary, advanced image analysis methods suitable for large-scale body composition analysis have been developed and evaluated. In future research the developed methods can be used, one at a time or in combination, for detailed studies of associations between body composition and, for example, prevalent and incident cardiometabolic diseases and their risk factors.