Author ORCID Identifier
Date of Graduation
Doctor of Philosophy (PhD)
Cervical cancer is one of the most common cancer in low- and middle-income countries (LMICs). The mortality rate can be reduced if radiation treatment becomes widely available. However, due to the lack of radiation treatment facilities and human resources, many cervical cancer patients in Africa are not able to receive timely treatments or advanced therapies. To increase the availability of radiation treatment in low-and middle-income countries (LMICs) including African countries, many attempts have been made to reduce the cost of medical linear accelerators. However, increasing the number of treatment machines would not instantly resolve the issues, as there would be insufficient trained and experienced medical staff to create high-quality radiation treatment plans. To fill the gap, we automated the entire radiation treatment planning process by automating the contouring, planning, and quality assurance (QA) processes in cervical cancer radiation treatment.
To create a high-quality radiation treatment plan, accurate contours must be generated first. We used convolutional neural networks (CNN), one of the most effective deep learning techniques for image processing, to create an auto-contouring model for 3 clinical target volumes (CTVs) and 12 normal structures for cervical cancer radiation treatment and showed that 93% of the automatically generated contours were clinically acceptable.
For planning, we automated 3 treatment delivery techniques including 2D 4-field-box, 3D conformal radiation therapy (3D-CRT), and volumetric-modulated arc therapy (VMAT). We also automated the field-in-field (FIF) technique to reduce hotspots in the automatically generated 4-field-box and 3D-CRT plans. Each beam delivery technique was evaluated on 35 retrospective patient datasets from South Africa, and on average, 95% of the automatically generated plans were clinically acceptable.
As clinically unacceptable plans were mostly caused by inaccurately generated contours, the quality of the contours should be verified to ensure the quality of the plans. To automatically detect clinically unacceptable contours, we developed an automated contour QA method using two independently developed auto-contouring systems. We hypothesized that if one of the two independently developed auto-contouring systems failed, the discrepancy between the two contours would be substantial enough to be identified by measuring the similarity between the two contours. We found that more than 90% of the contouring errors can be detected with an appropriate choice of similarity metrics.
In conclusion, the majority of the automatically generated contours and plans for cervical cancer radiation treatment were clinically acceptable. Furthermore, errors in the contours can be flagged by the contour QA method. The entire system has been implemented to the Radiation Planning Assistant (RPA), a web-based toolbox for automated planning, to help cervical cancer patients in LMICs.
radiotherapy, automated contouring, deep learning, automated treatment planning, low- and middle-income countries, cervical cancer, contour QA