Automatic delineation of hippocampus in CT images based on deep learning and dosimetry study in whole brain radiotherapy

Abstract

ObjectiveThe artificial intelligence algorithm based on deep learning, combined with multi-modal brain magnetic resonance imaging (MRI) and computed tomography (CT) images, was used to develop an automatic delineation technology of the hippocampal structure from CT images to provide an efficient and accurate automatic delineation method for hippocampal protection during cranial radiotherapy. MethodsThe MR and CT images of 20 patients with brain metastases were collected. After registering the MR and CT images, the hippocampus was delineated. Aunet, Unet, and Pix2pix deep learning models were trained on the CT-MRI dataset, and the differences between three automatically segmented hippocampus and the hippocampus manually segmented by senior chief physicians were calculated. The VMAT plan was designed separately for each patient. ResultsThe prediction results of the proposed method were more accurate and closer to the real segmentation results. The Dice, P, and R values were 0.8529, 0.8560, and 0.8632, respectively, indicating that the prediction results of Aunet were the closest to the real one. For the dose results, the differences between the maximum dose of the real hippocampus and three different methods were 75.2 (Aunet), 354.0 (Unet), and 462.1 (Pix2pix), and the differences between the average doses of the real hippocampus and three different methods were 30.3 (Aunet), 31.0 (Unet), and 66.6 (Pix2pix). ConclusionThe Aunet model can achieve efficient and accurate automatic delineation of the hippocampus on CT images. The dosimetric difference between the segmented and real hippocampi is extremely small, which is convenient to protect the hippocampus during brain radiotherapy.