Optimization of prostate cancer diagnosis using deep learning techniques to reduce false positives in PET with 18F-PSMA and 68Ga-PSMA (preliminary studies)

Prostate cancer is a leading cause of male mortality. The introduction of radiopharma- ceuticals like 68Ga-PSMA and F18-PSMA has significantly improved metastasis detection. However, F18-PSMA shows a 22% higher false positive rate than 68Ga-PSMA, highlighting the need for advanced tools to improve diagnostic accuracy. In this context, deep learning offers promising support in distinguishing true metastases from false positives. We tested nnUNet, a self-contouring deep learning model, to distinguish true metastases from false positives in F18-PSMA PET/CT scans. A dataset of 27 PET-CT scans (4 for testing) was used. nnUNet was chosen for its adaptability to various segmentation tasks, thanks to its au- tomated normalization and preprocessing pipeline. Training focused on auto-segmentation of metastatic lesions to reduce operator-dependent variability. Despite the limited dataset, nnUNet showed promising accuracy in segmenting lesions and reducing false positives. These findings suggest strong potential for AI in improving F18-PSMA PET imaging and support further validation on larger datasets.