Getting to the Heart of the Problem: Automated Image Segmentation of Cardiac Structures for Radiotherapy Dosimetry

Radiotherapy is a safe and effective treatment for many patients with cancer. However, incidental radiation exposure to healthy tissues can cause adverse side effects. In the treatment of thoracic tumours, such as for patients with breast or lung cancer, the heart is particularly at risk of exposure due to its proximity to the treatment fields. Long-term patient monitoring has provided clear evidence that thoracic radiotherapy can increase the risk of heart disease and cardiac-related mortality. While this risk is known to increase approximately linearly with the mean dose to the whole heart, more research is needed to better understand how the dose to different cardiac substructures is related to cardiotoxicity risks. Dosimetric analysis of large, retrospective patient datasets is required for further investigation. Between patients, anatomical variability and differences in radiotherapy techniques results in large variation in the doses to the cardiac chambers, valves, coronary arteries and other parts of the heart. For individualised cardiac substructure dosimetry, the different heart structures must be delineated on radiotherapy planning imaging. This presents a challenge: manual contouring is time-consuming, and inter-observer variability can produce substantial dosimetric uncertainty. A potential solution lies in automating the delineation process with image segmentation techniques. This thesis outlines the development, validation, and application of an automatic, atlas-based segmentation framework designed to delineate cardiac substructures for accurate and consistent dosimetry. Implementation of this framework across various radiotherapy datasets has demonstrated precise and robust substructure segmentation. For the whole heart volume, the resulting automatic segmentation achieved a Dice similarity coefficient of 0.944 ± 0.024, with a mean absolute surface-to-surface distance of 1.726 ± 1.363 mm. Accuracy in cardiac dose measurements was a primary consideration, with automatic segmentation of the whole heart resulting in an average difference in the mean heart dose of -0.20 ± 0.17 Gy (approximately 5-10%). Furthermore, incorporating novel developments has overcome known challenges to accurately delineate the coronary arteries, and provides detection and correction of errors so that the framework can be completely automated. Investigation into automatically estimating the dosimetric uncertainty associated with inter-observer contouring variability provides valuable information for both research and in a clinical setting. Finally, using this software in the analysis of a dataset of over 1500 Danish women treated for breast cancer highlighted the inadequacy of the mean whole heart dose to characterise individual patient cardiac dosimetry. The segmentation framework provides a clinically useful tool, a flexible research platform, and facilitates accurate dosimetry in large, retrospective studies. This underpins the much-needed development of detailed models of cardiotoxicity risks for patients receiving radiotherapy, aiding in the continued effort to improve patient outcomes.