A Dedicated Neural Network for Automated Segmentation of Prostate Gland from PET Images
Zeinab Shirkhani, Alireza Kamali‐Asl, Reza Jahangir, Hossein Arabi, Habib Zaidi · 2022
In PET imaging, prostate segmentation would provide useful quantitative information regarding the radiotracer absorption for radiomics studies with the aim of identifying the relationship between image characteristics and the patients’ conditions and outcomes. Since manual segmentation is a time-consuming process and prone to intra- and inter-observant errors/variations, development of robust and automated prostate segmentation is very crucial in clinical practice. The prostate segmentation from PET images is highly challenging due to the limited spatial resolution of the PET images as well as high noise levels normally present in the PET images. In this light, dedicated networks should be developed with meticulous training and hyper parameter optimization. To this end, a dedicated model based on UNet architecture has been proposed and evaluated in this study. The dedicated model would perform fully automated segmentation and identification of boundaries of prostate gland on PET images. 74 whole-body68Ga-PSMA PET/CT images were collected, wherein the ground truth masks for deep network training were generated through manual segmentation of PET images by an experienced nuclear medicine specialist. 52 and 22 subjects were randomly selected for model training and validation/test of the segmentation model. The model training was performed with and without data-augmentation techniques to examine their implication on prostate segmentation. The proposed model with morphological augmentation technique exhibited an accuracy of 0.881 in terms of Dice similarity coefficient (DSC) and 0.787 in terms of Intersection over Union (IoU) metric. The model developed with this augmentation technique led to no outlier in the segmentation of the prostate from PET images. Achieving a segmentation accuracy of 0.881 in terms of DSC metric indicated that the prostate segmentation from PET images is feasible with high accuracy.