Bone metastasis automatic segmentation from [18F]NaF whole-body PET using a convolutional neural network

Whole-body [
¹⁸
F]NaF positron emission tomography has attracted growing attention for its extremely sensitive bone imaging performance and outstanding diagnostic efficacy in identifying bone metastases during both the staging and restaging processes for patients with high-risk cancer. However, segmentation for whole-body imaging data is heavy and time-consuming work for radiologists and manual segmentation can leave out suspicious metastatic lesions. To utilize the automatic segmentation of suspicious bone metastatic lesions from [
¹⁸
F]NaF positron emission tomography, we explore convolutional neural network-based deep learning techniques to achieve an end-to-end segmentation method. The network was trained from the input [
¹⁸
F]NaF positron emission tomography standard uptake value image to produce a corresponding segmentation mask. The performance of the issued segmentation method was evaluated by typical segmentation criteria including precision, recall, F1-score and intersection over union. The quality of neural network-produced segmentation masks is good in general, and it showed the potential to assist clinicians and radiologists and simplify the process of segmentation substantially.