A computer-aided diagnosis tool in prostate cancer patients with biochemical recurrence using 18F-PSMA PET/CT imaging

Prostate cancer (PCa) patients frequently experience biochemical recurrence (BCR) following definitive primary treatment. Although fluorine-18-labeled prostate-specific membrane antigen positron emission tomography/computed tomography (^{1 8}F-PSMA PET/CT) is an imaging modality that is highly sensitive for BCR, false-positive findings owing to benign nonspecific uptake complicate diagnosis. Existing artificial intelligence (AI) tools have attempted to address this challenge but are often limited by their reliance on ground-truth labels derived from expert visual interpretation; thus, these tools reproduce expert opinion rather than confirming disease status.

The purpose of this study was to develop a computer-aided diagnosis (CAD) system for classifying benign and malignant findings on ^{1 8}F-PSMA PET/CT imaging in patients with PCa and BCR. Post-therapy imaging follow-up was used as an objective reference standard for malignancy.

A dataset of 69 patients with PCa and BCR who underwent ^{1 8}F-PSMA PET/CT imaging was used to develop a CAD. The system was evaluated under two classification schemes based on different ground truths: Task 1 used post-therapy imaging follow-up as the objective reference standard for malignant findings (in a subset of 45 patients), whereas task 2 relied on expert visual interpretation at baseline. In total, after data augmentation and filtering, 334 findings were analyzed for task 1, and 467 findings were analyzed for task 2. Suspicious findings were manually segmented using LifeX software (version 25.06.1). One-dimensional intensity profiles were extracted along the x-, y-, and z-axes at the maximum intensity voxel of each finding, from which profile-based and Gaussian-fit features were derived. The intensity profiles were used directly as inputs to a multilayer perceptron (MLP) classifier and were also used to extract profile- and Gaussian-fit-based features to train a random forest (RF) classifier using feature-importance analysis. The final model incorporated a stacking ensemble architecture combining the MLP and RF base models; logistic regression was the meta-classifier. The model was trained and evaluated using stratified 10-fold cross-validation. In each fold, 90% of the findings were assigned to the training set for model development, including feature selection, and the remaining 10% were held out as an independent test set for performance evaluation. Within each training set, five-fold internal cross-validation was used as the validation procedure for feature selection and stacking.

For task 1, the stacking ensemble model achieved an accuracy of 92.5% (SD, 3.5%), sensitivity of 92.3% (SD, 4.6%), specificity of 93.0% (SD, 9.2%), and an area under the receiver operating characteristic curve (AUROC) of 0.97. For task 2, performance was similar: accuracy, 91.7% (SD, 5.2%); sensitivity, 93.8% (SD, 4.5%); specificity, 85.8% (SD, 8.6%); and AUROC, 0.97. Feature-importance analysis revealed that raw intensity magnitude and spatial gradients along the x-axis were the most discriminative features for classification.

The proposed CAD system achieved highly accurate classification of ^{1 8}F-PSMA PET/CT findings, leveraging post-therapy imaging follow-up as a more objective reference standard for identifying malignancy than expert-based visual interpretation alone. The system achieved high and consistent performance across both the follow-up-based and expert-based labeling tasks. The integration of this tool into clinical workflow could improve diagnostic confidence and support the personalized management of PCa patients with BCR.