Confidence-based deep learning model for prostate cancer MRI boosts accuracy by rejecting uncertain predictions

In a study of 410 patients, a confidence-based AI model for prostate cancer on biparametric MRI retained 80.8% of predictions at a 0.8 threshold, achieving 94.6% accuracy vs. 90.8% for a senior radiologist, with 99 false-positives and 2 false-negatives.

• Before rejection, overall accuracy was 79.3%.
• Qualitative Grad-CAM visualization highlighted suspicious regions, but quantitative spatial agreement with lesion annotations was not assessed.
• The model was developed and tested on retrospective bpMRI data from 410 patients.

Automated summary

RadPigeon summaries are original and for information only. They are not clinical advice.