Introduction
Cardiac toxicity is a potential side-effect of thoracic radiotherapy, due to proximity of the heart to radiation tumour targets. However, there is currently no accurate way to quantify an individual’s post-treatment cardiac risk. Existing studies have focused on whole heart radiation dose, though there is increasing evidence that cardiac substructure doses may be more predictive of different detrimental effects1-4. Studying cardiac substructures is challenging, as manually delineating these small structures is labour-intensive, often resulting in small cohort numbers. There is thus a significant gap in our understanding of cardiac radiation dose limits, with no known “safe dose” to the heart.
Methods
Our multidisciplinary team set out to develop an automatic segmentation tool that can accurately and consistently delineate the heart and substructures. This tool can be utilised on thoracic radiotherapy planning CTs – including stereotactic ablative lung radiotherapy(SABR), non-SABR lung radiotherapy and breast radiotherapy - to examine radiation doses delivered to heart and substructures.
Results
We successfully developed and validated an automatic segmentation tool for 18 heart structures (whole heart, 4 chambers, 3 great vessels, 4 cardiac valves, 4 coronary arteries and 2 conduction nodes)5,6. This was achieved through combining the strengths of 3 different automatic segmentation techniques – deep learning, multi-atlas mapping, and geometric definitions for small hard-to-see structures. When compared with manual “gold standard” heart substructure delineations, the tool had a median accuracy of 2.1-8.6mm (mean-distance-to-agreement) and median variation in radiation dose of 0.9-6.8%. We successfully extended this tool to model uncertainties that occur during radiation planning, such as cardio-respiratory motion.
We have deployed this tool on a cohort of 117 SABR patients from an Australian phase 2 multi-centre trial (SSBROC-002). Analysis showed that for central tumours, the maximum dose can be as high as 51.7Gy to the heart. When stratified by the mean heart dose (MHD), Kaplan-Meier analysis suggest detrimental survival for the 50% who received higher than the median MHD (p=0.00004).
Conclusion
This novel and time-efficient automatic segmentation tool makes large-scale studies feasible, overcoming prior challenges associated with manual delineation. We demonstrated that it can be deployed on radiotherapy CT imaging, and current work involves analysing larger breast and lung cancer patient datasets. Further work also involves linking heart substructure doses with baseline cardiovascular risks, effects of systemic therapy and cardiac outcomes. This will enable development of an individualised cardiac risk-prediction model, with potential to reduce cardiac morbidity and mortality, leading to improved patient outcomes.