Glottic lesion segmentation of computed tomography images using deep learning

Abstrak

The larynx, a common site for head and neck cancers, is often overlooked in automated contouring due to its small size and anatomically complex nature. More than 75% of laryngeal tumors originate in the glottis. This paper proposes a method to automatically delineate the glottic tumors present contrast computed tomography (CT) images of the head and neck. A novel dataset of 340 images with glottic tumors was acquired and pre-processed, and a senior radiologist created a detailed, manual slice-by-slice tumor annotation. An efficient deep-learning architecture, the U-Net, was modified and trained on our novel dataset to segment the glottic tumor automatically. The tumor was then visualized with the corresponding ground truth. Using a combined metric of dice score and binary cross-entropy, we obtained an overlap of 86.68% for the train set and 82.67% for the test set. The results are comparable to the limited work done in this area. This paper’s novelty lies in the compiled dataset and impressive results obtained with the size of the data. Limited research has been done on the automated detection and diagnosis of laryngeal cancers. Automating the segmentation process while ensuring malignancies are not overlooked is essential to saving the clinician’s time.