Diagnostic performance of a fully automated AI algorithm for lesion detection and PI-RADS classification in patients with suspected prostate cancer

Abstract

Purpose: To evaluate the diagnostic performance of a fully automated, commercially available AI algorithm for detecting prostate cancer and classifying lesions according to PI-RADS. Material and methods: In this retrospective single-center cohort study, we included consecutive patients with suspected prostate cancer who underwent 3T MRI between May 2017 and May 2020. Histopathological ground truth was targeted transperineal ultrasound-fusion guided biopsy and extensive systematic biopsy. We compared the results of the AI algorithm to those of human readers on both the lesion and patient level and determined the diagnostic performance. Results: A total of 272 patients with 436 target lesions were evaluated. Of these patients, 135 (49.6%) had clinically significant prostate cancer (sPCa), 35 (12.9%) had clinically insignificant prostate cancer (ISUP = 1), and 102 (37.5%) were benign. On patient level, the cancer detection rates of sPCa for AI versus human readers were 11% versus 18% for PI-RADS ≤ 2, 27% versus 11% for PI-RADS 3, 54% versus 41% for PI-RADS 4, and 74% versus 92% for PI-RADS 5. The AI showed significantly higher accuracy: 74% versus 63% for PI-RADS ≥ 4 (p < 0.01) and 70% versus 52% for PI-RADS ≥ 3 (p < 0.01). Additionally, the AI correctly classified 62 patients with human reading PI-RADS ≥ 3 as true negatives. Conclusion: The AI algorithm proved to be a reliable and robust tool for lesion detection and classification. Its cancer detection rates and PI-RADS category distribution align with the results of recent meta-analyses, indicating precise risk stratification.