Two-Stage Approach for Segmenting Gross Tumor Volume in Head and Neck Cancer with CT and PET Imaging

Radiation treatment planning for head and neck cancers involves careful delineation of the tumor target volume on CT images, often with assistance from PET scans as well. In this study, we described a method to automatically segment the gross tumor volume of the primary tumor with a two-stage approach using deep convolutional neural networks as part of the HECKTOR challenge. We trained a classification network to select the axial slices which may contain the tumor, and these slices were then inputted into a segmentation network to generate a binary segmentation map. On the test set consisting of 53 patients, we achieved a mean Dice similarity coefficient of 0.644, mean precision of 0.694, and mean recall of 0.667.