Anatomic context-aware segmentation of organs-at-risk in thorax computed tomography scans

Abstract

Organ segmentation plays a crucial role in disease diagnosis and radiation therapy planning. Efficient and automated segmentation of the organs-at-risk (OARs) re-quires immediate attention since manual segmentation is a time consuming and costly task that is also prone to inter-observer variability. Automatic segmen-tation of organs-at-risk using deep learning is prone to predicting extraneous regions, especially in apical and basal slices of the organs where the shape is dif-ferent from the center slices. This thesis presents a novel method to incorporate prior knowledge on shape and anatomical context into deep-learning based organ segmentation. This prior knowledge is quantified using distance transforms that capture characteristics of the shape, location, and relation of the organ position with respect to the surrounding organs. In this thesis, the role of various distance transform maps has been explored to show that using distance transform regres-sion, alone or in conjunction with classification, improves the overall performance of the organ segmentation network. These maps can be the distance between each pixel and the center of the organ, or the closest distance between two organs; such as the esophagus and the spine. When used in a single-task regression model, these distance maps improved the segmentation results. Moreover, when used in a multi-task network with classification being the other task, they acted as regularizers for the classification task and yielded improved segmentations. The experiments were conducted on a computed tomography (CT) thorax dataset of 265 patients and the organs of interest are the heart, the esophagus, the lungs, and the spine. The results revealed a significant increase in f-scores and decrease in the Hausdorff distances for the OARs when segmented using the proposed model compared to the baseline network architectures.