Deep learning AI and Restriction Spectrum Imaging for patient-level detection of clinically significant prostate cancer on MRI
Yuze Song,Mariluz Rojo Domingo,Christopher Charles Conlin,Deondre D Do,Madison T Baxter,Anna Dornisch,George Xu,Aditya Bagrodia,Tristan Barrett,Mukesh Harisinghani,Gary Hollenberg,Sophia Kamran,Christopher J Kane,Dimitri A Kessler,Joshua Kuperman,Kang-Lung Lee,Michael A Liss,Daniel JA Margolis,Paul M Murphy,Nabih Nakrour,Truong Ngyuen,Thomas L Osinski,Rebecca Rakow-Penner,Shoumik Roychowdhury,Ahmed S Shabaik,Shaun Trecarten,Natasha Wehrli,Eric P Weinberg,Sean Woolen,Anders M Dale,Tyler M Seibert
DOI: https://doi.org/10.1101/2024.11.22.24317504
2024-11-26
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=.01) 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.
Radiology and Imaging