An artificial intelligence method for predicting postoperative urinary incontinence based on multiple anatomic...

Area of Science:

• Medical imaging and artificial intelligence
• Urology and oncology
• Machine learning interpretability

Background:

• Deep learning (DL) models offer potential in medicine but often lack interpretability.
• Captum, a tool for interpreting neural networks, is underutilized in medical research.
• Limited data exists on MRI anatomical measurements for prostate cancer patients post-Robotic-Assisted Radical Prostatectomy (RARP), hindering continence prediction.

Purpose of the Study:

• To explore the energy efficiency of DL models for predicting urinary continence after RARP using MRI data.
• To analyze and compare statistical and DL models for continence prediction.
• To provide a reference for applying interpretable DL models in clinical settings.

Main Methods:

• Utilized MRI anatomical measurements from patients who underwent RARP between July 2019 and December 2020.
• Employed statistical methods and computational models to identify continence features and evaluate their impact.
• Developed and analyzed a DL model (UINet7) for feature discovery and prediction.

Main Results:

• The UINet7 model achieved 0.97 accuracy in predicting continence.
• Age and six anatomical measurements were identified as the top seven continence features.
• The top four features identified by UINet7 were also confirmed by primary statistical analysis.

Conclusions:

• This study addresses the gap in understanding post-RARP continence features using interpretable DL.
• It presents a novel application of DL models to clinical urological problems.
• Interpretable AI analysis shows significant potential for clinical applications in predicting patient outcomes.