Deep Learning Artificial Intelligence and Restriction Spectrum Imaging for Patient-level Detection of Clinically...

Area of Science:

• Radiology and Medical Imaging
• Artificial Intelligence in Medicine
• Oncology

Background:

• Prostate cancer detection relies on MRI interpretation, often supplemented by scoring systems like PI-RADS.
• Advanced imaging techniques and AI offer potential improvements in diagnostic accuracy.
• The integration of Restriction Spectrum Imaging (RSI) and Deep Learning (DL) for csPCa detection requires evaluation.

Purpose of the Study:

• To assess if combining RSIrs_max and DL models with PI-RADS improves patient-level detection of clinically significant prostate cancer (csPCa).
• To compare the performance of RSIrs_max alone, DL models alone, and their combination with PI-RADS against PI-RADS alone.

Main Methods:

• A multi-institutional study included 1892 patients with biopsy-confirmed prostate cancer.
• Two DL models (3D-DenseNet and 3D-DenseNet+RSI) were developed and validated using biparametric MRI and RSI data.
• Performance was evaluated using ROC curves, AUC, sensitivity, and specificity, with comparisons made against PI-RADS interpretation.

Main Results:

• Neither RSIrs_max nor DL models alone significantly outperformed PI-RADS.
• Combining RSIrs_max with PI-RADS yielded an AUC of 0.78, while DL models + PI-RADS achieved an AUC of 0.80.
• Both combined approaches significantly improved csPCa detection compared to PI-RADS alone, with notable gains in specificity.

Conclusions:

• While RSIrs_max and DL models did not surpass PI-RADS individually, their combination with PI-RADS significantly enhances the detection of csPCa.
• The findings suggest that integrating advanced imaging data and AI with radiologist expertise can improve diagnostic performance for prostate cancer.
• Further research is needed to address limitations such as biopsy as a reference standard and external validation.