Artificial intelligence-based parotid contouring for radiation oncology in head and neck cancers.

Abstract

Background: Radiation therapy plays a critical role in head and neck cancer treatment, which can utilize artificial intelligence algorithms for automatic contour segmentation, treatment planning, and the selection of individual treatments. During the planning phase of radiation therapy, organs at risk (OARs), like the parotid gland, must be identified and mapped out to avoid unwanted side effects. Manual contouring of the parotid gland can be time-consuming and error-prone, causing interobserver variability. This study presents an algorithmic framework for the automated delineation of parotid glands using artificial intelligence and deep learning algorithms that interprets image sets to increase the precision and effectiveness of treatment planning, reduce the risk of side effects, and enhance treatment outcomes. Auto contouring increases the consistency and reproducibility of treatment plans and reduces the need for recontouring.

Methods: In this study, CT scans of 20 anonymized datasets were used. The slices were visualized with their corresponding contours. Further, the CT scans were preprocessed, and various image processing techniques were applied. The dataset generation process was automated, and the data were fed into the developed model. The U-Net architecture was used to create the model. The evaluation metrics used for the model were Accuracy, Precision, Recall, Loss, and Intersection Over Union (IOU).

Results: The proposed model gave a validation accuracy of auto contouring for the left contour to be ~97%. A validation accuracy for the right contour is ~96%.

Conclusions: The auto contouring of parotid glands proves to be a convenient and reliable framework that can be applied and used by radiation oncologists.