Verification of Automatic Detection of Prostate Cancer Lesion with68Ga-PSMA PET/CT Images Using Deep Supervised Residual U-Net

1353 Purpose: Characterizing lesions on PSMA PET/CT plays a critical role in the diagnosis, treatment planning and monitoring and theranostics of prostate cancer (PC). An end-to-end deep neural network has been developed to detect the PC lesions automatically based on data from European patients1. This study aims to verify the applicability of this method in Chinese patients scanned with a different scanner. Methods: 35 patients (mean age 70.3 ± 7.9 years, range 54-82 years) with pathologically confirmed prostate cancer was collected from PET center of Huashan Hospital, Fudan University. All patients underwent 68Ga-PSMA-11 PET/CT imaging using United Imaging uMI 510 PET/CT scanner (United Imaging, Shanghai, China) from head to thigh. The previously developed 3D deep supervised residual U-Net based on 193 European patients scanned with Siemens Biograph mCT PET/CT scanners (Siemens, Erlangen, Germany) was directly employed1. This network comprises a down-sampling path and an up-sampling path to learn the latent features. Following the previous study, the detection of lesions was focused in the pelvic area including the residual/relapse lesion, bone and lymph node metastasis. Results: For all lesions located in the pelvic area, the proposed method achieves precision of 94%, recall of 74.9%, and F1 score of 83.4%. The results are consistent with accuracy tested in European data. Conclusions: The preliminary experimental results demonstrated the feasibility of the developed AI method applying on heterogenous data with different origin. Extensive investigation of the robustness of AI algorithms is necessary before further development or clinical application.