Automatic Gross Tumor Volume Delineation of Nasopharyngeal Carcinoma in 3D CT Images.

Abstract

PURPOSE/OBJECTIVE(S)\nBased on a large-scale and multi-center dataset, we investigated the accuracy and efficiency of three-dimensional convolutional neural networks (3D-CNNs) for automatically delineating the gross tumor volume (GTV) of nasopharyngeal carcinoma (NPC). We further boost the performance by using self-supervision on whole-body scans to learn the general feature representation of normal tissue, by which CNN models will be able to further understand the recognition of lesions in CT scans.\n\n\nMATERIALS/METHODS\nWe constructed a multi-center dataset consisting of 170 CT scans of head and neck cases with nasopharyngeal cancer and 100 CT scans of the normal whole body scans. A 3D-CNN was developed to automatically predict the delineation. A pretext self-supervised learning strategy named context restoration, was used on whole-body scans to generate pre-trained weight. This design is helpful for CNN models to capture general knowledge about normal tissue such as organ shape, contexture, and so on, therefore leads to better GTV segmentation performance. Dice and IoU scores were calculated to evaluate the model s accuracy relative to ground truth labels.\n\n\nRESULTS\nThe pre-trained 3D-CNN achieved good performance with Dice of 73.1% (P-value < 0.025) and IoU of 58.9% (P-value < 0.025), on our collected multi-center dataset. Compared to conventional 2D models, this model has much superior performance (73.1% vs 66.3% on Dice). Compared to typical 3D model, this learning framework saved about 75% training time to achieve the same performance on training set consisting of 100 CT scans (reducing the training from two days to eleven hours), and improves the segmentation performance by about 8% on Dice value. 10 CT scans without cancer are extensively used to evaluate the ability of addressing normal tissue and it s found that using pretraining strategy let the false detection area decrease by a large margin (false detected scan number decreased from 8 to 2).\n\n\nCONCLUSION\nThe proposed 3D-CNN was capable of automatically delineating the gross tumor volume for nasopharyngeal cancer and outperforms traditional 2D schemes. The AI system costs nearly 75% less time to build compared to networks used in previous works, which is more suitable in clinics.