Deep convolutional neural network for automatic segmentation and classification of jaw tumors in contrast-enhanced computed tomography images.

Abstract

The purpose of this study was to evaluate the performance of convolutional neural network (CNN)-based image segmentation models for segmentation and classification of benign and malignant jaw tumors in contrast-enhanced computed tomography (CT) images. A dataset comprising 3416 CT images (1163 showing benign jaw tumors, 1253 showing malignant jaw tumors, and 1000 without pathological lesions) was obtained retrospectively from a cancer hospital and two regional hospitals in Thailand; the images were from 150 patients presenting with jaw tumors between 2016 and 2020. U-Net and Mask R-CNN image segmentation models were adopted. U-Net and Mask R-CNN were trained to distinguish between benign and malignant jaw tumors and to segment jaw tumors to identify their boundaries in CT images. The performance of each model in segmenting the jaw tumors in the CT images was evaluated on a test dataset. All models yielded high accuracy, with a Dice coefficient of 0.90-0.98 and Jaccard index of 0.82-0.97 for segmentation, and an area under the precision-recall curve of 0.63-0.85 for the classification of benign and malignant jaw tumors. In conclusion, CNN-based segmentation models demonstrated high potential for automated segmentation and classification of jaw tumors in contrast-enhanced CT images.