Radiology

Deep-learning CAD for screening prostate MRI shows good discrimination but lower specificity than radiologists

In a prostate cancer screening cohort of 1254 men, 1354 biparametric MRI exams were used to train and test an nnU-Net–based AI system. Clinically significant prostate cancer was defined as ISUP grade ≥2, with histopathology and ≥3-year follow-up as reference standards. The AI achieved an AUROC of 0.83 (95% CI 0.73–0.92) for detecting clinically significant cancer. At matched sensitivity, AI specificity was significantly lower than that of radiologists using PI-RADS v2 criteria. 1

CT radiomics models outperform net water uptake for classifying stroke onset before versus after 4.5 hours

This study analyzed 993 ischemic stroke patients with known onset-to-imaging times from the MR CLEAN Registry and LATE trial. Machine learning models used clinical-radiological data, net water uptake, lesion volume, and radiomics features from non-contrast CT. Models based on clinical scores or net water uptake alone achieved AUCs of 0.65 for classifying onset within versus beyond 4.5 hours. Combining net water uptake with lesion volume or clinical variables improved performance to AUCs of 0.70–0.75. 2

PSMA-PET staging improves PSA persistence and 36‑month BCR-free survival in high-risk cN0M0 prostate cancer

This multi-institutional study included 1475 high-risk prostate cancer patients undergoing radical prostatectomy and extended pelvic lymph node dissection. Patients staged with PSMA-PET (miN0M0) were compared with those staged by conventional imaging (cN0M0) using 1:1 propensity score matching. After matching, PSA persistence occurred in 3.2% of PSMA-PET patients versus 14% of conventionally staged patients (p < 0.001). Thirty-six–month biochemical recurrence–free survival was 90.9% in the PSMA-PET group versus 82.2% with conventional imaging. 3