A discriminative learning based approach for automated nasopharyngeal carcinoma segmentation leveraging multi-modality similarity metric learning
Zongqing Ma, Xi Wu, Shanhui Sun, Chaoyang Xia, Zhipeng Yang, Shuo Li, Jiliu Zhou · 2018
The combination of imaging information from multi-modality images may be highly beneficial for radiotherapy treatment planning in terms of tumor delineation. This paper proposes a discriminative learning based approach for automated nasopharyngeal carcinoma (NPC) segmentation using multi-modality images. Specially, an image-patch-based multi-modality convolutional neural network (CNN) is designed to jointly learn a multi-modality similarity metric and classification of paired image patch of different modalities. The CNN integrates two normal classification sub-networks into a Siamese-like sub-network. With the help of the multi-modality similarity metric learning provided by the Siamese-like sub-network, the classification sub-networks are able to take advantage of each other's multimodal information. Validation of our method was performed on 50 CT-MR subjects. Experimental results demonstrate our method achieves improved segmentation performance compared to its counterpart without multi-modality similarity metric learning and the segmentation method of solely using CT, with a Dice Similarity Coefficient metric of 0.712 compared to 0.659 and 0.636.