Purpose: Accurate localizing and segmenting tumors is important for diagnosis and treatment in head & neck (H&N) cancer. We aim to develop a hybrid neural network (HNN) model, fully automatic tumor localization and segmentation through a combination of Faster-RCNN / U-Net in CT images.
Methods: In this preliminary study, a total of 48 patients with PET and CT images in H&N obtained from The Cancer Imaging Archive are used. It consists of complete CT/PET and Gross Tumor Volume (GTV) contours, and 825 slices were selected for training, 176 for validation and 79 for testing. HNN consists of two phases in the training stage, they are Faster-RCNN for tumor localization and U-Net for tumor segmentation. Each phase is trained independently. Since PET can detect the tumor location very well and Faster-RCNN is an outstanding model for learning object localization PET is fed into Faster-RCNN to learn the tumor localization. Furthermore, since U-Net has been proved an excellent medical image segmentation model, it is used for tumor segmentation in CT images. The testing stage consists of two steps: (1) tumor localization and segmentation: the PET image is fed into the trained Faster-RCNN which produces a class label and bounding box coordinates, these coordinates are used to automatically crop the testing CT image to a fixed-sized window, which is then fed into the U-Net for segmentation; (2) OTSU threshold method is applied for post-processing. The segmentation results were evaluated using accuracy, sensitivity, and specificity.
Results: The proposed model achieved the average values of 0.959, 0.930, and 0.962 for accuracy, sensitivity, and specificity respectively.
Conclusion: In this preliminary study, a hybrid neural network (HNN) for tumor localization and segmentation was proposed in H&N cancer. The experimental results demonstrated that HNN can localize tumors very well and obtain promising segmentation performance as well.