UCSF

Prostate cancer is a significant global health concern, ranking as the second most commonlydiagnosed cancer and fifth leading cause of cancer-related deaths in men worldwide. Prostatespecific membrane antigen (PSMA)-targeted positron emission tomography (PET) is used for staging, especially in intermediate to high-risk cases and biochemical recurrence, offering superior sensitivity and specificity compared to conventional methods. Despite advancements, the analysis of PSMA-PET data remains largely manual, prompting the need for computer-aided diagnosis using machine learning or deep learning to enhance efficiency and consistency. Some previous work have been done to develop a prediction model based on deep neural network (DNN), however it failed in prediction of high-volume disease cases. In this project, I hypothesized that refined lesion annotation and specialized batching strategies can enhance algorithm performance, leading to improved segmentation accuracy, and SUV measurements. The successful implantation of better contour of the lesion was done by a clinical fixed-threshold method. The results suggested the refined lesion annotations significantly impacted the SUVmean values in lesions and will further affected the result of DNN training.