Artificial Intelligence for Breast Cancer Radiotherapy: Automated Contouring, Dose–Volume Prediction, and Clinical Workflow Evaluation in Bangladesh
Authors: M. Monir Uddin
Institution: North South University
Introduction
Radiotherapy for breast cancer requires accurate outlining of the treatment targets and nearby organs at risk (OARs), followed by careful evaluation of the planned radiation dose. These steps can take considerable time and require experienced staff, which can be challenging in busy radiotherapy centers. The workload can be particularly difficult in low- and middle-income countries (LMICs), where the demand for cancer treatment is increasing while specialist resources remain limited. Artificial intelligence (AI) may help reduce this workload by assisting with routine parts of radiotherapy planning. We developed an AI system to automatically outline treatment targets and OARs and to predict important dose–volume parameters for breast cancer radiotherapy. We evaluated the system using data from four radiotherapy centers in Bangladesh and assessed its usefulness in the clinical workflow with radiation oncologists.
Methods
We retrospectively collected more than 1,000 breast cancer radiotherapy cases from four institutions in Bangladesh: Bangladesh Medical University, Square Hospital, Labaid Hospital, and United Hospital. The AI system used patient imaging and radiotherapy planning information to automatically generate contours of treatment targets and clinically important OARs and to predict dose–volume parameters. The accuracy of the generated contours was compared with expert clinical contours using standard segmentation measures. The predicted dose–volume parameters were compared with those from the actual clinical treatment plans. We also performed a retrospective multi-institutional reader study involving nine radiation oncologists with different levels of experience. They assessed the AI-assisted workflow using measures such as task completion time and other clinically relevant performance measures. The system was also implemented in a web-based interface that allows radiation oncologists to review the AI- generated contours and dose-related information.
Results
The AI system achieved mean Dice similarity scores above 90% for the evaluated treatment targets and OARs, indicating close agreement between the AI-generated and expert clinical contours. The predicted dose–volume parameters also showed good agreement with the corresponding measurements from the clinical treatment plans. In the multi-institutional reader study, the AI-assisted workflow reduced the time required for the evaluated radiotherapy planning tasks while maintaining clinically acceptable performance. The benefit was observed among radiation oncologists with different levels of experience. A web-based system was developed to allow clinicians to review the AI-generated contours and dose-related information and incorporate them into the radiotherapy planning workflow.
Conclusion
Our findings suggest that AI can assist with breast cancer radiotherapy planning using real- world clinical data collected from multiple centers in Bangladesh. By combining automated contouring with dose–volume prediction and clinical workflow evaluation, the system may help reduce the workload associated with radiotherapy planning while keeping radiation oncologists responsible for the final treatment decisions. The web-based implementation provides a practical platform for further clinical evaluation and may be particularly useful in radiotherapy centers where specialist time and resources are limited.
