AUTHOR=Wang Tonghe , Lei Yang , Schreibmann Eduard , Roper Justin , Liu Tian , Schuster David M. , Jani Ashesh B. , Yang Xiaofeng 

TITLE=Lesion segmentation on 18F-fluciclovine PET/CT images using deep learning

JOURNAL=Frontiers in Oncology

VOLUME=Volume 13 - 2023

YEAR=2023

URL=https://www.frontiersin.org/journals/oncology/articles/10.3389/fonc.2023.1274803

DOI=10.3389/fonc.2023.1274803

ISSN=2234-943X

ABSTRACT=18F-fluciclovine (anti-3-18 F-FACBC), has been demonstrated to be associated with significantly improved survival when used in PET/CT imaging to guide postprostatectomy salvage radiotherapy . We aim to investigate the feasibility of using deep learning method to automatically detect and segment lesions on 18F-fluciclovine PET/CT images. We retrospectively identified 84 patients enrolled in the Arm B of Emory Molecular Prostate Imaging for Radiotherapy Enhancement (EMPIRE-1) trial. All the 84 patients had prostate adenocarcinoma, underwent prostatectomy and 18F-fluciclovine PET/CT imaging with lesions identified and delineated by physicians. Three different neural networks with increasing complexities, U-net, Cascaded U-net, and cascaded detection segmentation network, were trained and tested on the 84 patients with a five-fold cross validation strategy and a hold-out test using manual contours as ground truth. We also investigated using both PET and CT, or using PET only as input of the neural network. Dice similarity coefficient (DSC), 95th percentile Hausdorff distance (HD95), center-of-mass distance (CMD) and volume difference (VD) were used to quantify the quality of segmentation results against ground truth contours. All the three deep learning methods can detect 144/155 lesions and 153/155 lesions successfully when PET+CT and PET only serving as input, respectively. Quantitative results demonstrated that the neural network with the best performance can segment lesions with an average DSC of 0.68±0.15 and HD95 of 4±2 mm. The center-of-mass of the segmented contours deviates from physician contours by around 2 mm on average, and the volume difference is less than 1cc. The novel network proposed by us achieves the best performance compared to current networks. The addition of CT as input into the neural network contributed to more failing cases (DSC=0), and among those cases of DSC>0, it is shown to have no statistically significant difference when compared with using only PET as input for our proposed method.: Quantitative results demonstrated the feasibility of the deep learning methods in automatically segmenting lesions on 18F-fluciclovine PET/CT images. It indicates the great potential of 18F-fluciclovine PET/CT combining with deep learning by providing a second check in identifying lesions as well as saving time and effort for physicians in contouring.