Deep learning models built from PSMA PET of the primary tumor can predict synchronous and metachronous prostate cancer metastases

Abstract

Objective The objective was to develop prognostic models that included convolutional neural networks (CNN) derived from 18 F-DCFPyL (PSMA) PET imaging of the primary tumor uptake patterns to prognose early metastatic progression after curative intent treatment for localized prostate cancer. Methods Due to the lack of sufficient cases with adequate follow-up and metastatic events to derive this model directly, we derived models that predict the presence of synchronous metastases using only data obtained from the primary tumor. Because early metastatic progression events are consequent to occult metastases present at the time of initial therapy, we hypothesized that a model trained to predict synchronous metastases might also predict metachronous metastatic progression. A convolutional neural network (CNN) model was generated using whole-prostate PSMA PET images and auto-segmented intraprostatic lesions and combined with clinicopathologic data and imaging parameters to develop a multimodal model. Model performance was evaluated using the area under the receiver operating characteristic curve (AUC) for predicting synchronous metastases and metachronous metastatic progression. Results The multimodal model and CNN model had AUCs of 0.82 (95% CI 0.69–0.92, p < 0.005) and 0.72 (95% CI 0.55–0.84, p 0.0059), respectively, for prediction of synchronous metastases. Shapley additive explanation analysis showed the CNN had the largest contribution to the combined model performance. For metastatic progression, the multimodal model had an AUC of 0.839 (95% CI: 0.6763–1.000, p = 0.0064). Conclusion The multimodal model trained to predict synchronous metastases also predicted metachronous metastatic progression. This supports the potential of artificial intelligence applied to primary tumor PSMA PET images for enhanced prognostication. However, significant limitations include the modest sample size, single center source of data, and potential for overfitting, which may limit generalizability. Therefore, further validation in a larger cohort is indicated.