Post-Prostatectomy Risk Stratification of Biochemical Recurrence Using Transfer Learning-Based Multi-Modal Artificial Intelligence

Abstract

For patients undergoing radical prostatectomy for prostate cancer (PCa), accurate risk stratification is essential to guide post-prostatectomy therapeutic decision making. Recently, there has been success in the use of multi-modal artificial intelligence models for men after prostate biopsy to aid in risk stratification. Herein, we trained and tested a TRansfer learning-based multi-modal Artificial InteLligence model (TRAIL) for biochemical recurrence (BCR) risk stratification following radical prostatectomy. Patients contained within a prospective PCa registry at a single institution were utilized. Digital pathology slides from the diagnostic biopsies prior to radical prostatectomy for patients with clinically localized PCa were scanned at 20x resolution. Features were extracted for the TRAIL model from pathology slides via two transfer learning steps: (1) InceptionResNetv2 that first determines a heatmap of tumor areas, and (2) A ResNet18 that extracts representative features from the high tumor probability areas. Least Absolute Shrinkage and Selection Operator (LASSO) was used for feature selection from the pathology-extracted features. Finally, TRAIL combines the clinical and pathology-extracted features via a classification ensemble model based on weak tree learners to predict 2- and 5-year BCR defined as two consecutive serum PSA levels ≥0.2 ng/mL. TRAIL training was performed on 250 patients and was then locked and applied to the test set of 125 patients. Accuracy and the area under the curve (AUC) were calculated. Comparison to CAPRA-S and to clinical-only features were assessed. A total of 818 digital whole pathology biopsy slides from 375 patients treated with subsequent radical prostatectomy were included. Surgical margins were positive in 29% of the patients, and 41% had extra-prostatic extension. The median follow-up was 48 months (Range: 1-132 months). The rates of 2-and 5-year BCR were 11% and 18% respectively. A total of 19 digital pathology-driven features were included in TRAIL. Clinical factors included age, ISUPG, Gleason score, PSA, pathological T and N stages, surgical margin involvement, and the presence of extra-prostatic extension. On the testing set, TRAIL achieved a 2-year BCR AUC of 0.76 and accuracy of 0.87, and was superior to CAPRA-S (AUC = 0.57) and clinical-only features (AUC 0.50, accuracy 0.14). For 5-year BCR, TRAIL achieved an AUC of 0.69 and accuracy of 0.78, and performed better than CAPRA-S (AUC = 0.58), and clinical only features (AUC = 0.50, accuracy = 0.23). Through a combination of deep and ensemble learning, TRAIL incorporates clinical and histopathology features, enabling an improved BCR risk stratification post-prostatectomy when compared to the currently used clinicopathologic models. Future work with larger datasets with metastatic events is warranted to further optimize the model for clinical use.