The study utilized a clinical dataset of 71 male patients who underwent PET/CT imaging. For each patient, the Maximum Intensity Projection (MIP) was created on the PET coronal plane series. Manual cropping of regions of interest (ROIs) was performed on suspicious findings, based on the diagnoses of experienced nuclear medicine physicians, generating 157 ROIs in total, of which 76 were labeled as FP and 81 as TP. Each ROI image was resized to 224x224.

Then, the images were divided into two datasets: original and augmented, and individually processed using the DenseNet201 model. The augmented dataset was generated using transformations such as rotations, shifts, and flips via ImageDataGenerator, resulting in more than 12.000 images dynamically created during training. The model was optimized using 512 units with L2 regularization over 100 epochs. The datasets were split into training and testing using an 80:20 ratio.  Using five-fold cross-validation, accuracies, specificities, sensitivities, and the Area Under the ROC Curves (AUROCCs) were calculated.