Development of a deep learning-based system for aiding in the determination of Prostate Imaging Reporting and Data System (PI-RADS) scores: an international multicenter study

Background. Prostate multiparametric magnetic resonance imaging is widely recommended prior to biopsy in clinical practice, with the Prostate Imaging Reporting and Data System (PI-RADS) as the standard tool for guiding diagnosis and treatment decisions. However, analyzing multiparametric magnetic resonance imaging data demands substantial expertise, and the process is often time-intensive and cognitively challenging, leading to variability between and within readers.Aim. To create a deep learning-based computer-aided diagnosis (DL-CAD) system to minimize manual influence on PI-RADS score determination.Materials and methods. Between January 2020 and May 2024, 108 patients with histopathologically confirmed prostate cancer with PI-RADS scores 4–5 were retrospectively selected for model development and training. Additionally, 28 benign cases were included for model validation. Different prostate zones were labeled following PI-RADS v2.1 guidelines to facilitate model selection. Manual segmentation of prostate regions and lesions was performed on T2-weighted (T2W) sequences, and a 3D U-Net architecture was implemented for the DL model using the MONAI framework. Diagnostic performance was assessed using Python-based statistical analysis.Results. The DL-CAD system achieved average accuracy of 78 %, sensitivity of 60 %, and specificity of 84 % for lesion detection. The Dice similarity coefficient for prostate segmentation was 0.71, and the AUROC was 81.16 %.Conclusion. The DL-CAD system demonstrates promise for patients with clinically significant prostate cancer by improving diagnostic accuracy. While it exhibits high specificity, further improvements of sensitivity and segmentation accuracy are necessary. These improvements could be achieved through the use of larger datasets and advanced deep learning techniques, such as transfer learning or ensemble learning, which could enhance sensitivity without compromising specificity. Further multicenter validation is required to accelerate the integration of this system into clinical practice.