AUTOMATED DEEP LEARNING SEGMENTATION OF HEAD AND NECK TUMORS AND ORGANS-AT-RISK FOR RADIOTHERAPY PLANNING

Abstract

This work proposes a deep learning automated segmentation system to precisely localize head and neck tumors and critical organs-at-risk (OAR) for radiotherapy planning. The proposed approach will aim to improve the segmentation accuracy, reduce the clinical workload, and improve the treatment planning efficiency. A set of computed tomography (CT) and magnetic resonance imaging (MRI) scans of large number of patients with head and neck cancer was utilized. The image was preprocessed by normalization, augmentation and artifact reduction techniques. A novel segmentation architecture based on convolutional neural networks ( CNN) with multi-scale feature extraction, attention mechanism and encoder-decoder pathways has been designed and successfully implemented to accurately segment multiple OARs and gross tumour volumes (GTVs). Four metrics were used to evaluate the performance of the model: Dice Similarity Coefficient (DSC), Intersection over Union (IoU), Hausdorff Distance (HD), precision, recall and volumetric overlap.