Magnetic resonance imaging (MRI) is the primary modality for diagnosing prostate cancer, with the Prostate Imaging–Reporting and Data System (PI-RADS) used for assessment, grading lesions from 1 to 5. However, its diagnostic accuracy remains suboptimal (PI-RADS 4: 59% [39–78%]; PI-RADS 5: 85% [73–94%]). Currently, no superior imaging interpretation method is available. This study aims to enhance prostate cancer diagnosis by reconstructing actual pathological slides from prostate specimens, correlating them with preoperative MRI, and employing artificial intelligence (AI) models for training and automated identification. The ultimate goal is to develop a more accurate MRI-based diagnostic model beyond the traditional PI-RADS system. In the future, this approach could be integrated with robotic-assisted biopsy systems, improving diagnostic accuracy and minimizing the need for repeated biopsies, thereby reducing the risk of sepsis and bleeding.

A retrospective analysis was conducted on patients diagnosed with prostate cancer who underwent radical prostatectomy at Taichung Veterans General Hospital between January 2007 and July 2024. Patients without digitized whole-mount pathology slides were excluded. Postoperative pathological lesions were correlated with preoperative MRI findings to evaluate the diagnostic performance of the conventional PI-RADS system. Lesions with different Gleason scores were separately marked and compared with the corresponding MRI images to identify any distinguishing imaging features associated with different Gleason grades. The labeled pathological slides and their corresponding MRI images were ultimately utilized to train an artificial intelligence model for detecting prostate cancer lesions on MRI. Subsequently, we assessed whether the AI-based model could enhance diagnostic accuracy.

Between January 2007 and July 2024, we identified 126 patients with preoperative prostate MRI scans and whole-mount prostate pathology confirming prostate cancer. The average number of whole-mount slides per patient was 6.12 (range: 4–8), varying based on prostate size and lesion location. Among them, 99 patients (78%) had radiology reports indicating suspicious prostate lesions on MRI. By correlating lesions identified on whole-mount pathology with MRI findings, we observed a mismatch rate of 17.7%. A higher mismatch rate was associated with pathological slides showing a greater number of lesions, fragmented or distorted specimens, or longer intervals between the MRI and surgery.

We observed a 17.7% mismatch rate in the localization of prostate cancer lesions between whole-mount pathology and MRI scans. This discrepancy was associated with a higher number of lesions, fragmented or distorted specimens, and longer intervals between MRI and surgery. Moving forward, we aim to train AI models using these labeled pathological slides and corresponding MRI scans, and to evaluate the diagnostic accuracy of the AI-assisted approach.