Evaluation of an artificial intelligence model based on multiparametric transrectal ultrasound for localizing clinically significant prostate cancer by simulation of targeted biopsies.
Authors
Affiliations (11)
Affiliations (11)
- Department of Urology, Amsterdam UMC, Amsterdam, The Netherlands. [email protected].
- Cancer Centre Amsterdam, Amsterdam, The Netherlands. [email protected].
- Lab of Biomedical Diagnostics, Department of Electrical Engineering, Eindhoven University of Technology, Eindhoven, The Netherlands.
- Department of Urology, Amsterdam UMC, Amsterdam, The Netherlands.
- Department of Urology, Leiden University Medical Center, Leiden, The Netherlands.
- Cancer Centre Amsterdam, Amsterdam, The Netherlands.
- Julius Center for Health Sciences and Primary Care, UMC Utrecht, Utrecht University, Utrecht, The Netherlands.
- Angiogenesis Analytics, Den Bosch, The Netherlands.
- Department of Urology, Canisius Wilhelmina Hospital, Nijmegen, The Netherlands.
- Prosper Collaborative Prostate Cancer Clinics, Nijmegen-Eindhoven, The Netherlands.
- Department of Urology, Netherlands Cancer Institute - Antoni van Leeuwenhoek Hospital, Amsterdam, The Netherlands.
Abstract
An AI model that performs well during training does not guarantee similar performance in clinical practice and should be carefully evaluated before implementation. We aimed to evaluate a voxel-level trained AI model (AUROC 0.87), which utilizes a three-dimensional multiparametric transrectal prostate ultrasound (3D mpUS) to identify clinically significant prostate cancer (csPCa). We included patients with csPCa (Grade Group ≥ 2 and scheduled for radical prostatectomy (RP)) and without csPCa (PI-RADS ≤ 2 and/or negative systematic biopsies). Histopathology of RP specimens provided the csPCa reference standard. 3D mpUS consisted of grayscale, contrast-enhanced ultrasound, and shear-wave elastography using automated acquisition. We assessed patient-level diagnostic accuracy by comparing the results of simulated targeted biopsies based on the AI model with the reference standard in internal and external evaluation. Patients without csPCa and RP reference standard were used to determine specificity. Based on internal evaluation of 250 patients, a sensitivity of 0.82 (CI 0.75 to 0.87) and specificity of 0.43 (CI 0.32 to 0.55) was reached for ISUP ≥ 2. For ISUP ≥ 3, this was 0.90 (CI 0.83-0.95) and 0.39 (CI 0.31-0.47). In the external evaluation of 77 patients, the sensitivity for ISUP ≥ 2 was 0.81 (CI 0.65-0.90), with a specificity of 0.42 (CI 0.28-0.57). For ISUP ≥ 3, this was 0.96 (CI 0.78-0.99) and 0.42 (CI 0.30-0.55). The AI model based on 3D mpUS showed consistent patient-level performance for csPCa detection in internal and external evaluation, comparable to voxel-level analysis. These suggest strong generalizability and support prospective clinical trials. NCT04605276. Question Does the diagnostic performance of a 3D multiparametric ultrasound-based AI model translate from voxel-level training to patient-level biopsy simulation? Findings Simulated biopsy performance aligned with voxel-level results, showing robust csPCa detection and supporting the model's generalizability across independent datasets. Clinical relevance The AI model's consistent biopsy simulation performance confirms its readiness for clinical evaluation and suggests diagnostic value in MRI-constrained settings.