Deep learning AI and Restriction Spectrum Imaging for patient-level detection of clinically significant prostate cancer on MRI

Abstract Background The Prostate Imaging Reporting & Data System (PI-RADS), based on multiparametric MRI (mpMRI), is widely used for the detection of clinically significant prostate cancer (csPCa, Gleason Grade Group (GG≥2)). However, its diagnostic accuracy can be impacted by variability in interpretation. Restriction Spectrum Imaging (RSI), an advanced diffusion-weighted technique, offers a standardized, quantitative approach for detecting csPCa, potentially enhancing diagnostic consistency and performing comparably to expert-level assessments. Purpose To evaluate whether combining maximum RSI-derived restriction scores (RSIrs-max) with deep learning (DL) models can enhance patient-level detection of csPCa compared to using PI-RADS or RSIrs-max alone. Materials and Methods Data from 1,892 patients across seven institutions were analyzed, selected based on MRI results and biopsy-confirmed diagnoses. Two deep learning architectures, 3D-DenseNet and 3D-DenseNet+RSI (incorporating RSIrs-max), were developed and trained using biparametric MRI (bpMRI) and RSI data across two data splits. Model performance was compared using the area under the receiver operating characteristic curve (AUC) for patient-level csPCa detection, using PI-RADS performance for clinical reference. Results Neither RSIrs-max nor the best DL model combined with RSIrs-max 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 0.79 (95% CI: 0.74-0.83; P =.005) for combination of RSIrs-max with PI-RADS and 0.81 (95% CI: 0.76-0.85; P <.001) for combination of best DL model with PI-RADS, compared to 0.73 (95% CI: 0.68-0.78) for PI-RADS alone. Conclusion Both RSIrs-max and DL models demonstrate comparable performance to PI-RADS alone. Integrating either model with PI-RADS significantly enhances patient-level detection of csPCa compared to using PI-RADS alone. Summary Statement RSIrs-max and deep learning models match the performance of expert PI-RADS in patient-level csPCa detection and combining either with PI-RADS yields a significant improvement over PI-RADS alone. Key Points In a study of 1,892 patients from seven institutions undergoing MRI and biopsy for prostate cancer, RSIrs-max and the DL model (AUC, 0.75 ( P =.59) and 0.78 ( P =.09)) performed comparably to expert-level PI-RADS scores (AUC, 0.73). Including prostate auto-segmentation improved the DL model (AUC, 0.68 ( P =.01) vs 0.72 ( P =.60)). Combining RSIrs-max or the DL model (AUC, 0.79 ( P =.005) and 0.81 ( P <.001)) with PI-RADS statistically significantly outperformed PI-RADS alone (AUC, 0.73).

Fetched live from OpenAlex and de-inverted. Abstracts are not stored in this database: the inverted indexes are 8.6 GB of the frame’s 9.3 GB of text, and the host has 13 GB free.