Automated Contouring of Pelvic Computed Tomography Images for Radiation Therapy Systems using Deep Learning

Accurate delineation of anatomical regions of interest on computed tomography images is an important step in radiotherapy planning for linear accelerator-based treatments associated with cancer. However, manual contouring is time consuming, labor intensive, and associated with inter-observer variability, which can prove to be a challenge in resource-congested environments. This paper suggests an automated contouring system implemented with deep learning to overcome these shortcomings, and with special focus on its practical application in developing nations. To achieve the desired computational complexity, a custom Mini U-Net architecture was created and acceptable contouring accuracy was achieved. To be clinically relevant and usable, the model was incorporated into a DICOM compliant, clinician in the loop graphical user interface, which allowed a smooth interface with the existing radiotherapy workflows. The proposed framework has been assessed on the basis of its results against the results of pelvic CT images of females who were used, and in which different ROI areas were used to calculate the results. During testing, it highlighted satisfactory outcomes of different areas with a Dice score of 0.98 for the outline of the patient area; however, other areas were less accurate in computing the outcomes. In this case, the rectum area achieved a Dice score of merely 0.14 due to insufficient training materials. However, overall, the results have shown the efficacy and potential use of the proposed contouring system for significant reductions in manual contouring times without a compromise in segmentation accuracy and low computational complexity for the considered ROIs. This work also shows the effectiveness and prospects of using DL techniques for contouring in radiotherapy and the areas to improve, such as availability of data for certain anatomical structures.