Emerging research suggests that multimodal artificial intelligence (AI) may bolster risk stratification in predicting post-prostatectomy biochemical recurrence (BCR) of prostate cancer (PCa).
For the retrospective study, recently published in Clinical Imaging, researchers developed and compared a multimodal deep learning model to the Cancer of the Prostate Risk Assessment post-Surgical score (CAPRA-S) for predicting BCR in 311 patients who had multiparametric magnetic resonance imaging (mpMRI) prior to undergoing a radical prostatectomy. The multimodal AI model (identified as M4 in the study) incorporated MRI-based AI features and the clinical variables of age and prostate-specific antigen (PSA) level, according to the study.
The study authors found that the fully automated multimodal deep learning model offered a 74 percent area under the receiver operating characteristic curve (AUROC) in contrast to a 68 percent AUROC for CAPRS-S scores > 6 in the testing cohort. While the deep learning model offered 20 percent lower specificity than CAPRS-S scores > 6 (66 percent vs. 86 percent), it provided nearly double the sensitivity (75 percent vs. 38 percent), according to the researchers.
“Although this increased sensitivity is at the cost of lower specificity, sensitivity can be more important for flagging higher-risk patients for close postsurgical monitoring, considering BCR is most effectively treated when caught earlier. It is possible that automated deep learning-based features could usefully be integrated into the current medical workflow after future validation and equity analysis,” wrote lead study author Benjamin D. Simon, M.D., who is affiliated with the Molecular Imaging Branch at the National Cancer Institute and the National Institutes of Health in Bethesda, Md., and colleagues.
In a subgroup analysis of patients deemed to have intermediate risk for BCR, the M4 deep learning model offered a 23 percent higher AUROC in contrast to CAPRA-S (74 percent vs. 51 percent).
Three Key Takeaways
- Improved sensitivity for predicting BCR. The multimodal AI model (M4) demonstrated substantially higher sensitivity (75 percent) than the traditional CAPRA-S score >6 (38 percent) for predicting post-prostatectomy biochemical recurrence (BCR), which may help in earlier identification of high-risk patients.
- Enhanced risk stratification in intermediate-risk patients. In patients with intermediate BCR risk, the M4 model achieved a notably higher AUROC (74 percent vs. 51 percent) compared to CAPRA-S, offering improved prognostic discrimination for a clinically challenging subgroup.
- Multimodal AI integration shows prognostic value. By combining deep learning-derived mpMRI features with clinical variables (age and PSA), the M4 model showed evidence of survival discrimination in Kaplan-Meier analysis, suggesting potential utility for personalized post-surgical monitoring strategies.
In addition to providing the highest sensitivity of the evaluated deep learning models in this cohort (60 percent), the researchers said log-rank testing revealed that the M4 deep learning model was the only one to demonstrate statistical significance in differentiating between those with BCR and patients who were BCR-free.
“ … Our M4 model extends prognostic capability beyond imaging alone by integrating deep learning-derived quantitative imaging features with relevant clinical parameters, achieving some evidence of predictive discrimination compared to PI-RADS based on the Kaplan-Meier analysis although there was no statistically significant difference in AUROC,” added Simon and colleagues. “Notably, our Kaplan-Meier analysis indicated robust survival discrimination, especially in clinically challenging intermediate-risk groups, suggesting a substantial improvement in personalized prognostic stratification.”
(Editor’s note: For related content, see “MRI-Based Deep Learning Model Bolsters Prediction of PI-RADS 3 and $ Lesions,” “Adjunctive AI Bolsters Lesion-Level PPVs for csPCa in International bpMRI Study” and “Study: Adjunctive AI Provides Over 18 Percent Higher Lesion-Level Sensitivity on Prostate MRI.”)
Beyond the inherent limitations of a single-center retrospective study, the authors acknowledged the use of endorectal coils for many patients and the lack of assessment for time to BCR.
