A Method for Prostate Cancer Metastases Classification In 18f-PSMA PET/CT Images Using a Machine Learning Stacking Ensemble Model
Abstract
Purpose
To develop a computer-aided diagnosis (CAD) system for the classification of benign versus malignant findings on 18F-PSMA PET/CT imaging. It was applied in patients with biochemical recurrent (BCr) prostate cancer (PCa), using post-therapy follow-up (FU) imaging information (tumor regression or lesion regression in size indicated malignancy) as the Gold Standard.
Methods
A dataset of 44 patients with BCr PCa who underwent 18F-PSMA PET/CT imaging was analyzed. Suspicious findings were categorized based on post-therapy imaging FU. Findings that did not respond to therapy and were interpreted as non-specific by the physician were labeled as benign, whereas findings that demonstrated treatment response were labeled as malignant. A total of 285 findings were included (Benign: 123, Malignant: 162), with data augmentation applied on the benign class to mitigate class imbalance. For each finding, the maximum intensity point was selected, and 1D intensity profiles along the x, y, and z axes were extracted in the axial plane, directly used as inputs to a multi-layer perceptron (MLP) classifier. These intensity profiles were also used to extract statistical and Gaussian-fit features to train a random forest (RF) classifier. The final model includes a stacking ensemble architecture combining MLP and RF base models with logistic regression serving as the meta-classifier. Forward feature selection was used, and the model was trained with 10-fold cross validation.
Results
Our model achieved a mean accuracy of 92.6±5.3%, sensitivity of 94.4±6.2%, specificity of 90.3±10.7%, and Area Under the Receiver Operating Characteristic Curve (AUROC) of 97.4±3.6%. Feature selection analysis showed that the top 15 features ranked by importance were predominantly profile-derived intensity features in contrast to Gaussian fit–based features.
Conclusion
Using post-therapy imaging FU, this work demonstrates that lesion-level intensity profile–based features can differentiate benign from malignant findings on 18F-PSMA PET/CT images, highlighting the importance of FU validated CAD approaches.