# Deep Learning Artificial Intelligence and Restriction Spectrum Imaging for Patient-level Detection of Clinically Significant Prostate Cancer on Biparametric Magnetic Resonance Imaging

## TL;DR

This study shows that combining AI models or advanced imaging data with radiologist assessments improves detection of aggressive prostate cancer in MRI scans.

## Contribution

The novel contribution is demonstrating that integrating AI or restriction spectrum imaging scores with radiologist evaluations enhances patient-level cancer detection.

## Key findings

- Neither restriction scores nor AI models alone outperformed radiologist assessments.
- Combining AI or restriction scores with radiologist evaluations significantly improved cancer detection.
- The approach increased specificity without reducing sensitivity for aggressive prostate cancer.

## Abstract

Combining restriction scores derived from advanced restriction spectrum imaging (RSIrs) with deep learning (DL) offers a complementary advantage to expert Prostate Imaging-Reporting and Data System (PI-RADS) assessment. While neither the maximum RSIrs nor a DL model alone surpassed radiologist PI-RADS performance, integration of either approach with PI-RADS significantly improved patient-level detection of clinically significant prostate cancer. These findings suggest that quantitative imaging and artificial intelligence tools can enhance, rather than replace, expert interpretation in magnetic resonance imaging for prostate cancer diagnosis.

Our aim was to evaluate whether combining the maximum restriction score derived from restriction spectrum imaging (RSIrsmax) with deep learning (DL) models can enhance patient-level detection of clinically significant prostate cancer (csPCa) in comparison to Prostate Imaging-Reporting and Data System (PI-RADS) or RSIrsmax alone.

A total of 1892 patients from seven institutions who underwent imaging between January 2016 and March 2024 were included on the basis of magnetic resonance imaging (MRI) findings and biopsy-confirmed prostate cancer diagnosis. Two DL architectures, 3D-DenseNet and 3D-DenseNet+RSI (incorporating RSIrsmax), were developed and trained using biparametric MRI and RSI data using a leave-one-center-out validation approach. RSI is a rapid sequence that requires only 2–3 min to acquire. Model performance was evaluated in a biopsy-confirmed subset of 876 patients, with subgroup analyses stratified by site and scanner vendor. Receiver operating characteristic (ROC) and precision recall curves and forest plots (I2 for heterogeneity) were generated, and the area under the ROC curve (AUC) and sensitivity, were compared, as well as specificity at fixed sensitivity of 0.90. Calibration, decision-curve, and reclassification analyses (net reclassification improvement and integrated discrimination improvement) were performed. Codes used in developing the DL model are available on GitHub (https://github.com/ESONG1999/Deep-learning-AI-and-RSI-for-patient-level-detection-of-csPCa-on-MRI).

Neither RSIrsmax nor the best DL model combined with RSIrsmax significantly outperformed PI-RADS interpretation by expert radiologists. However, when combined with PI-RADS, both approaches significantly improved patient-level csPCa detection, with AUCs of 0.78 (95% confidence interval [CI] 0.75–0.81; p < 0.001) for RSIrsmax + PI-RADS and 0.80 (95% CI 0.77–0.82; p < 0.001) for the best DL model + PI-RADS, versus 0.75 (95% CI 0.71–0.78) for PI-RADS alone. The absolute gain in specificity at fixed sensitivity of 0.90 was 0.04 (95% CI 0.04–0.04) for RSIrsmax + PI-RADS, and 0.03 (95% CI 0.03–0.04) for DL + PI-RADS.

Both RSIrsmax and the best DL model demonstrated comparable performance to PI-RADS alone. Addition of either model to PI-RADS significantly enhanced patient-level detection of csPCa in comparison to PI-RADS alone. Limitations include biopsy as an imperfect reference, the exclusion of hip implant cases, lack of external calibration, limited RSI availability, and missing case-level information for individual radiologists and their expertise.

We looked at whether adding advanced scan data (ASD) and artificial intelligence (AI) models to radiologist assessments of MRI (magnetic resonance imaging) scans was better in detecting aggressive prostate cancer (PCa). We found that adding AI models or ASD to standard scan scores improved cancer detection in comparison to standard scores alone. The results suggest that combining radiologist expertise with AI and ASD may help in earlier identification of more patients with csPCa.