Evaluating the Predictive Value of Preoperative MRI for Erectile Dysfunction Following Radical Prostatectomy

Accurate preoperative prediction of erectile dysfunction (ED) is important for counseling patients undergoing radical prostatectomy. While clinical features are established predictors, the added value of preoperative MRI remains underexplored. We investigate whether MRI provides additional predictive value for ED at 12 months post-surgery, evaluating four modeling strategies: (1) a clinical-only baseline, representing current state-of-the-art; (2) classical models using handcrafted anatomical features derived from MRI; (3) deep learning models trained directly on MRI slices; and (4) multimodal fusion of imaging and clinical inputs. Imaging-based models (maximum AUC 0.569) slightly outperformed handcrafted anatomical approaches (AUC 0.554) but fell short of the clinical baseline (AUC 0.663). Fusion models offered marginal gains (AUC 0.586) but did not exceed clinical-only performance. SHAP analysis confirmed that clinical features contributed most to predictive performance. Saliency maps from the best-performing imaging model suggested a predominant focus on anatomically plausible regions, such as the prostate and neurovascular bundles. While MRI-based models did not improve predictive performance over clinical features, our findings suggest that they try to capture patterns in relevant anatomical structures and may complement clinical predictors in future multimodal approaches.

💡 Research Summary

This paper investigates whether pre‑operative multiparametric MRI can add predictive value for post‑radical prostatectomy erectile dysfunction (ED) at 12 months, beyond established clinical risk factors. Using a retrospective cohort of 647 prostate‑cancer patients from the Netherlands Cancer Institute, the authors built and compared four modeling strategies: (1) a clinical‑only baseline employing traditional machine‑learning algorithms on pre‑operative variables (age, BMI, smoking/alcohol use, comorbidities, baseline IIEF‑15 score, etc.); (2) handcrafted MRI features derived from automated segmentation of the prostate and the surrounding fascia, including regional fascia thickness (12 radial sectors per slice, median per sector) and volumetric estimates across 12 axial slices; (3) end‑to‑end deep‑learning (DL) models trained directly on 2‑D MRI slices (ResNet‑18, Vision Transformer‑B/16, and a hybrid ResNet‑ViT), evaluated with different slice configurations (single mid‑prostate slice, 4 mid‑slices, 8 base + mid slices, 12 consecutive slices); and (4) an intermediate multimodal fusion model that jointly learns embeddings from the best‑performing DL architecture and the clinical variables, concatenating them before a final classifier. A nested stratified 5‑fold outer cross‑validation with 3‑fold inner hyper‑parameter tuning (Optuna) ensured robust performance estimates; missing clinical data led to exclusion of 139 patients from the clinical and fusion experiments.

Results: The clinical‑only Random Forest achieved the highest discrimination (AUC 0.663, balanced accuracy 0.612, F1 0.58). Handcrafted MRI features performed only marginally better than chance (best AUC 0.554, SVM), indicating limited independent signal. DL models modestly outperformed handcrafted features, with the hybrid ResNet‑ViT using four mid‑prostate slices attaining the best AUC 0.569 and balanced accuracy 0.595. Adding more slices generally improved robustness, but inclusion of base slices did not help, likely due to lower image quality and higher anatomical variability. The multimodal intermediate‑fusion model yielded a slight improvement over imaging‑only (AUC 0.586, balanced accuracy 0.603) but still fell short of the clinical baseline. SHAP analysis on the fusion model showed clinical features contributed 57 % of the total importance (age, weight, baseline erectile function score being top), while imaging contributed 43 %. Saliency maps from the best DL model highlighted the prostate gland and adjacent neurovascular bundles, confirming that the network focuses on anatomically plausible regions.

Segmentation quality: nnU‑Net trained on 124 manually annotated cases achieved Dice scores of 0.965 ± 0.014 for the prostate and 0.738 ± 0.162 for the fascia on mid‑prostate slices, with lower performance at the apex. This enabled automated extraction of the handcrafted features but also revealed the inherent difficulty of delineating the thin fascia.

The authors conclude that, with the current dataset size, 2‑D imaging approach, and heterogeneity of MRI acquisition protocols, pre‑operative MRI does not provide independent predictive power beyond established clinical variables for post‑operative ED. Nevertheless, the fact that DL models attend to relevant anatomical structures suggests that MRI could become a useful complementary modality in larger, more standardized cohorts, especially if 3‑D volumetric data and precise neurovascular bundle annotations become available. Future work should explore domain‑adaptation techniques, balanced loss functions for class imbalance, and richer multimodal architectures to fully exploit the potential synergy between imaging and clinical data.