Lung cancer remains one of the leading causes of cancer-related mortality worldwide, with radiation therapy being a cornerstone of treatment. However, radiation-induced cardiotoxicity has emerged as a significant long-term complication, impacting patient survival and quality of life. There is increasing evidence that post-radiotherapy cardiotoxicity is better indicated by cardiac substructure dose, compared with whole heart dose. Despite this, cardiac substructures are not routinely incorporated into clinical workflows due to the significant time and resource demands associated with manual segmentation. Moreover, the relationship between dose distribution in radiotherapy plans and radiation-induced cardiotoxicity remains incompletely understood. Therefore, the development of an automated and precise segmentation approach for cardiac substructures, alongside a substructure-based cardiotoxicity prediction model, is essential for enabling the clinical application of cardiac substructure-specific dose constraints to mitigate cardiotoxicity risk. This study aims to establish automated tools to facilitate personalized cardiotoxicity management strategies in lung cancer radiotherapy.
First, a deep learning-based auto-segmentation model was developed to delineate 19 cardiac substructures on non-contrast planning computed tomography (CT) scans. The performance of the nnU-Net model was compared with the widely utilized 3D U-Net architecture. Subjective evaluation by four physicians determined that 94% of the automatically generated contours were clinically acceptable. Additionally, the same deep learning framework was employed to develop an auto-segmentation model for cardiac chambers on daily magnetic resonance (MR) images acquired with optimized flip angles. This model was specifically designed for adaptive radiotherapy workflows on MR-Linac systems, achieving a clinical acceptability rate of 95%. These models provide accurate and efficient cardiac substructure segmentation, facilitating both retrospective dosimetric analyses and prospective clinical applications.
Next, the relationship between cardiac substructure dose exposure and a cardiotoxicity biomarker, high-sensitivity cardiac troponin T (hs-cTnT) was investigated to investigate the role of individual cardiac substructures in radiation-induced toxicity. A rigorous analytical framework was implemented to assess the predictive value of clinical factors, dosimetric parameters, and radiomic features derived from both whole-heart and substructure-based metrics using two independent datasets. The findings demonstrated that cardiac substructure dosimetric parameters exhibited superior predictive performance, with left anterior descending coronary artery (LAD) V20Gy identified as the most significant predictor of hs-cTnT elevation.
Finally, a cardiotoxicity risk estimation and radiotherapy plan re-optimization framework was developed by integrating the auto-segmentation and toxicity prediction models. Objective functions were formulated based on recommended substructure dose thresholds and dynamically adjusted during the optimization process. The re-optimized treatment plans achieved significant reductions in LAD V20Gy and right ventricle maximum dose while maintaining adequate target coverage and adherence to organ-at-risk constraints.
In conclusion, this study establishes an artificial intelligence (AI)-assisted framework for cardiotoxicity management in lung cancer radiotherapy, incorporating automated segmentation models for both CT and MR imaging, a cardiotoxicity prediction model, and an optimization strategy for treatment planning. These tools substantially enhance the feasibility of cardiotoxicity risk assessment and mitigation, improve the understanding of radiation-induced cardiotoxicity, and enable patient-specific treatment plan modifications for improved clinical outcomes.
","abstract_html":"<p>Lung cancer remains one of the leading causes of cancer-related mortality worldwide, with radiation therapy being a cornerstone of treatment. However, radiation-induced cardiotoxicity has emerged as a significant long-term complication, impacting patient survival and quality of life. There is increasing evidence that post-radiotherapy cardiotoxicity is better indicated by cardiac substructure dose, compared with whole heart dose. Despite this, cardiac substructures are not routinely incorporated into clinical workflows due to the significant time and resource demands associated with manual segmentation. Moreover, the relationship between dose distribution in radiotherapy plans and radiation-induced cardiotoxicity remains incompletely understood. Therefore, the development of an automated and precise segmentation approach for cardiac substructures, alongside a substructure-based cardiotoxicity prediction model, is essential for enabling the clinical application of cardiac substructure-specific dose constraints to mitigate cardiotoxicity risk. This study aims to establish automated tools to facilitate personalized cardiotoxicity management strategies in lung cancer radiotherapy.</p> <p>First, a deep learning-based auto-segmentation model was developed to delineate 19 cardiac substructures on non-contrast planning computed tomography (CT) scans. The performance of the nnU-Net model was compared with the widely utilized 3D U-Net architecture. Subjective evaluation by four physicians determined that 94% of the automatically generated contours were clinically acceptable. Additionally, the same deep learning framework was employed to develop an auto-segmentation model for cardiac chambers on daily magnetic resonance (MR) images acquired with optimized flip angles. This model was specifically designed for adaptive radiotherapy workflows on MR-Linac systems, achieving a clinical acceptability rate of 95%. These models provide accurate and efficient cardiac substructure segmentation, facilitating both retrospective dosimetric analyses and prospective clinical applications.</p> <p>Next, the relationship between cardiac substructure dose exposure and a cardiotoxicity biomarker, high-sensitivity cardiac troponin T (hs-cTnT) was investigated to investigate the role of individual cardiac substructures in radiation-induced toxicity. A rigorous analytical framework was implemented to assess the predictive value of clinical factors, dosimetric parameters, and radiomic features derived from both whole-heart and substructure-based metrics using two independent datasets. The findings demonstrated that cardiac substructure dosimetric parameters exhibited superior predictive performance, with left anterior descending coronary artery (LAD) V20Gy identified as the most significant predictor of hs-cTnT elevation.</p> <p>Finally, a cardiotoxicity risk estimation and radiotherapy plan re-optimization framework was developed by integrating the auto-segmentation and toxicity prediction models. Objective functions were formulated based on recommended substructure dose thresholds and dynamically adjusted during the optimization process. The re-optimized treatment plans achieved significant reductions in LAD V20Gy and right ventricle maximum dose while maintaining adequate target coverage and adherence to organ-at-risk constraints.</p> <p>In conclusion, this study establishes an artificial intelligence (AI)-assisted framework for cardiotoxicity management in lung cancer radiotherapy, incorporating automated segmentation models for both CT and MR imaging, a cardiotoxicity prediction model, and an optimization strategy for treatment planning. These tools substantially enhance the feasibility of cardiotoxicity risk assessment and mitigation, improve the understanding of radiation-induced cardiotoxicity, and enable patient-specific treatment plan modifications for improved clinical outcomes. </p>","abstract_has_math":false,"creators":["Chen, Xinru","Lung cancer remains one of the leading causes of cancer-related mortality worldwide, with radiation therapy being a cornerstone of treatment. However, radiation-induced cardiotoxicity has emerged as a significant long-term complication, impacting patient survival and quality of life. There is increasing evidence that post-radiotherapy cardiotoxicity is better indicated by cardiac substructure dose, compared with whole heart dose. Despite this, cardiac substructures are not routinely incorporated into clinical workflows due to the significant time and resource demands associated with manual segmentation. Moreover, the relationship between dose distribution in radiotherapy plans and radiation-induced cardiotoxicity remains incompletely understood. Therefore, the development of an automated and precise segmentation approach for cardiac substructures, alongside a substructure-based cardiotoxicity prediction model, is essential for enabling the clinical application of cardiac substructure-specific dose constraints to mitigate cardiotoxicity risk. This study aims to establish automated tools to facilitate personalized cardiotoxicity management strategies in lung cancer radiotherapy.
First, a deep learning-based auto-segmentation model was developed to delineate 19 cardiac substructures on non-contrast planning computed tomography (CT) scans. The performance of the nnU-Net model was compared with the widely utilized 3D U-Net architecture. Subjective evaluation by four physicians determined that 94% of the automatically generated contours were clinically acceptable. Additionally, the same deep learning framework was employed to develop an auto-segmentation model for cardiac chambers on daily magnetic resonance (MR) images acquired with optimized flip angles. This model was specifically designed for adaptive radiotherapy workflows on MR-Linac systems, achieving a clinical acceptability rate of 95%. These models provide accurate and efficient cardiac substructure segmentation, facilitating both retrospective dosimetric analyses and prospective clinical applications.
Next, the relationship between cardiac substructure dose exposure and a cardiotoxicity biomarker, high-sensitivity cardiac troponin T (hs-cTnT) was investigated to investigate the role of individual cardiac substructures in radiation-induced toxicity. A rigorous analytical framework was implemented to assess the predictive value of clinical factors, dosimetric parameters, and radiomic features derived from both whole-heart and substructure-based metrics using two independent datasets. The findings demonstrated that cardiac substructure dosimetric parameters exhibited superior predictive performance, with left anterior descending coronary artery (LAD) V20Gy identified as the most significant predictor of hs-cTnT elevation.
Finally, a cardiotoxicity risk estimation and radiotherapy plan re-optimization framework was developed by integrating the auto-segmentation and toxicity prediction models. Objective functions were formulated based on recommended substructure dose thresholds and dynamically adjusted during the optimization process. The re-optimized treatment plans achieved significant reductions in LAD V20Gy and right ventricle maximum dose while maintaining adequate target coverage and adherence to organ-at-risk constraints.
In conclusion, this study establishes an artificial intelligence (AI)-assisted framework for cardiotoxicity management in lung cancer radiotherapy, incorporating automated segmentation models for both CT and MR imaging, a cardiotoxicity prediction model, and an optimization strategy for treatment planning. These tools substantially enhance the feasibility of cardiotoxicity risk assessment and mitigation, improve the understanding of radiation-induced cardiotoxicity, and enable patient-specific treatment plan modifications for improved clinical outcomes.
"]},{"key":"dc:title","label":"Title","values":["Artificial Intelligence (AI)-assisted Cardiotoxicity Management in Lung Cancer Radiotherapy"]}]}],"canonical_facts":{"dc:contributor":["Jinzhong Yang","Zhongxing Liao","Laurence E. Court"],"dc:creator":["Chen, Xinru","Lung cancer remains one of the leading causes of cancer-related mortality worldwide, with radiation therapy being a cornerstone of treatment. However, radiation-induced cardiotoxicity has emerged as a significant long-term complication, impacting patient survival and quality of life. There is increasing evidence that post-radiotherapy cardiotoxicity is better indicated by cardiac substructure dose, compared with whole heart dose. Despite this, cardiac substructures are not routinely incorporated into clinical workflows due to the significant time and resource demands associated with manual segmentation. Moreover, the relationship between dose distribution in radiotherapy plans and radiation-induced cardiotoxicity remains incompletely understood. Therefore, the development of an automated and precise segmentation approach for cardiac substructures, alongside a substructure-based cardiotoxicity prediction model, is essential for enabling the clinical application of cardiac substructure-specific dose constraints to mitigate cardiotoxicity risk. This study aims to establish automated tools to facilitate personalized cardiotoxicity management strategies in lung cancer radiotherapy.
First, a deep learning-based auto-segmentation model was developed to delineate 19 cardiac substructures on non-contrast planning computed tomography (CT) scans. The performance of the nnU-Net model was compared with the widely utilized 3D U-Net architecture. Subjective evaluation by four physicians determined that 94% of the automatically generated contours were clinically acceptable. Additionally, the same deep learning framework was employed to develop an auto-segmentation model for cardiac chambers on daily magnetic resonance (MR) images acquired with optimized flip angles. This model was specifically designed for adaptive radiotherapy workflows on MR-Linac systems, achieving a clinical acceptability rate of 95%. These models provide accurate and efficient cardiac substructure segmentation, facilitating both retrospective dosimetric analyses and prospective clinical applications.
Next, the relationship between cardiac substructure dose exposure and a cardiotoxicity biomarker, high-sensitivity cardiac troponin T (hs-cTnT) was investigated to investigate the role of individual cardiac substructures in radiation-induced toxicity. A rigorous analytical framework was implemented to assess the predictive value of clinical factors, dosimetric parameters, and radiomic features derived from both whole-heart and substructure-based metrics using two independent datasets. The findings demonstrated that cardiac substructure dosimetric parameters exhibited superior predictive performance, with left anterior descending coronary artery (LAD) V20Gy identified as the most significant predictor of hs-cTnT elevation.
Finally, a cardiotoxicity risk estimation and radiotherapy plan re-optimization framework was developed by integrating the auto-segmentation and toxicity prediction models. Objective functions were formulated based on recommended substructure dose thresholds and dynamically adjusted during the optimization process. The re-optimized treatment plans achieved significant reductions in LAD V20Gy and right ventricle maximum dose while maintaining adequate target coverage and adherence to organ-at-risk constraints.
In conclusion, this study establishes an artificial intelligence (AI)-assisted framework for cardiotoxicity management in lung cancer radiotherapy, incorporating automated segmentation models for both CT and MR imaging, a cardiotoxicity prediction model, and an optimization strategy for treatment planning. These tools substantially enhance the feasibility of cardiotoxicity risk assessment and mitigation, improve the understanding of radiation-induced cardiotoxicity, and enable patient-specific treatment plan modifications for improved clinical outcomes.
"],"dc:identifier":["https://digitalcommons.library.tmc.edu/utgsbs_dissertations/1434"],"dc:subject":["Cardiac toxicity; Segmentation; Cardiac substructure; Outcome prediction; Personalized treatment; Plan optimization","Other Medical Sciences"],"dc:title":["Artificial Intelligence (AI)-assisted Cardiotoxicity Management in Lung Cancer Radiotherapy"],"thesis:degree_level":["Dissertation (PhD)"],"thesis:degree_name":["Doctor of Philosophy (PhD)"]},"updated_at":"2026-07-24T05:49:08Z"}