ESTRO 2024 - Abstract Book

S3066

Physics - Autosegmentation

ESTRO 2024

1 Tygerberg Hospital, Medical Imaging and Clinical Oncology, Cape Town, South Africa. 2 Stellenbosch University, Medical Physics, Stellenbosch, South Africa

Purpose/Objective:

Cervical cancer is an ever-growing burden in Sub-Saharan Africa (SSA) and currently ranks the second highest in cancer incidence in SSA.1 Image guided High Dose Rate (IG-HDR) intracavitary brachytherapy is an irreplaceable curative treatment modality for locally advanced cervical cancer. Although IG-HDR brachytherapy comes with great advantages, the treatment planning workflow as a whole is a labour-intensive process that includes several manual, time-consuming steps and requires input from a range of professionals; Radiation Oncologists (ROs), Radiotherapists (RTs) and Medical Physicists (MPs). These issues are amplified in low- and middle-income countries, such as those SSA, as they face several challenges in radiotherapy. These challenges include but are not limited to high workloads, equipment maintenance, as well as significant deficits in experienced radiation oncologists, radiotherapists, and medical physicists.2,3 A recent study into the current state of cancer in sub-Saharan Africa (SSA), published in the Lancet Oncology journal, found that the shortages of radiotherapy professionals is one of the most crucial barriers hindering access to cancer services in SSA, with an estimated 211 % increase in workforce required to provide equitable access to radiotherapy.1 Several studies have applied machine learning (ML) based automation to tackle a variety of areas in radiotherapy, such as improving clinical workflows, quality assurance, chart reviews, treatment planning as well as organ at risk and target delineation.4,5 These automation tools can be of great benefit to countries in SSA, as they have the ability to substantially reduce bottlenecks in the treatment planning process. Therefore, this work focused on training and evaluating the novel self-configuring deep convolution neural network package, known as nnU-Net, for the automatic delineation of the organs at risk (OARs) and high-risk clinical target volume (HR CTV) for Ring-Tandem (R-T) IG-HDR cervical brachytherapy treatment planning. The computed tomography (CT) scans of 100 previously treated patients were used to train and test three different nnU-Net configurations (2D, 3DFR and 3DCasc). Of the 100 patients, 80 were used for training and validation while the remaining 20 patients were used for testing. The best configuration was determined by evaluating the performance of each model based on the calculated Sørensen-Dice similarity coefficient (DSC), Hausdorff distance (HD), 95th percentile Hausdorff distance (HD95), mean surface distance (MSD) and precision score for the 20 test patients. The dosimetric accuracy between the manual and predicted contours (generated from the best performing model) was assessed by looking at the various dose volume histogram (DVH) parameters and volume differences. The clinical acceptability of the generated contours was evaluated by three different radiation oncologists who scored the predicted bladder, rectum and HR CTV contours on a scale of 1 to 4 where; 1 indicated reject completely, 2 major revisions, 3 minor revisions before clinically acceptable and 4 indicated clinically acceptable as is. Contours scored with a 4, also included ones where the radiation oncologist would have preferred some small adjustments but were not necessary for the contours to be deemed clinically acceptable. For any contours scoring a 3 or lower, the radiation oncologists indicated the time required to adjust the contours to the level of clinical acceptability. These times, along with the model prediction times were compared to the average time taken for ROs to manually contour all the OARs and HR-CTV. Material/Methods:

Results: