Русский
Development of a Deep Learning-Based System for Supporting Medical Decision-Making in PI-RADS Score Determination
DOI: https://dx.doi.org/10.18565/urology.2024.6.5-11
He Mingze, Enikeev M.E., Rzaev R.T., Chernenkiy I.M., Feldsherov M.V., Li He, Hu Kebang, Shpot E.V., Glybochko P.V.
1) Institute for Urology and Reproductive Health, Sechenov University, Moscow, Russia; 2) Department of Radiology, The Second University Hospital, Sechenov University, Moscow, Russia; 3) Department of Radiology, The First Hospital of Jilin University, Changchun, China; 4) Department of Urology, The First Hospital of Jilin University, Changchun, China
Aim: to explore the development of a computer-aided diagnosis (CAD) system based on deep learning (DL) neural networks aimed at minimizing human error in PI-RADS grading and supporting medical decision-making.
Materials and Methods. This retrospective multicenter study included a cohort of 136 patients, comprising 108 cases of PCa (PI-RADS score 4–5) and 28 cases of benign conditions (PI-RADS score 1–2). The 3D U-Net architecture was applied to process T2-weighted images (T2W), diffusion-weighted images (DWI), and dynamic contrast-enhanced images (DCE). Statistical analysis was conducted using Python libraries to assess diagnostic performance, including sensitivity, specificity, Dice similarity coefficients, and the area under the receiver operating characteristic curve (AUC).
Results. The DL-CAD system achieved an average accuracy of 78%, sensitivity of 60%, and specificity of 84% for detecting lesions in the prostate. The Dice similarity coefficient for prostate segmentation was 0.71, and the AUC was 81.16%. The system demonstrated high specificity in reducing false-positive results, which, after further optimization, could help minimize unnecessary biopsies and overtreatment.
Conclusion. The DL-CAD system shows potential in supporting clinical decision-making for patients with clinically significant PCa by improving diagnostic accuracy, particularly in minimizing intra- and inter-observer variability. Despite its high specificity, improvements in sensitivity and segmentation accuracy are needed, which could be achieved by using larger datasets and advanced deep learning techniques. Further multicenter validation is required for accelerated integration of this system into clinical practice.
Materials and Methods. This retrospective multicenter study included a cohort of 136 patients, comprising 108 cases of PCa (PI-RADS score 4–5) and 28 cases of benign conditions (PI-RADS score 1–2). The 3D U-Net architecture was applied to process T2-weighted images (T2W), diffusion-weighted images (DWI), and dynamic contrast-enhanced images (DCE). Statistical analysis was conducted using Python libraries to assess diagnostic performance, including sensitivity, specificity, Dice similarity coefficients, and the area under the receiver operating characteristic curve (AUC).
Results. The DL-CAD system achieved an average accuracy of 78%, sensitivity of 60%, and specificity of 84% for detecting lesions in the prostate. The Dice similarity coefficient for prostate segmentation was 0.71, and the AUC was 81.16%. The system demonstrated high specificity in reducing false-positive results, which, after further optimization, could help minimize unnecessary biopsies and overtreatment.
Conclusion. The DL-CAD system shows potential in supporting clinical decision-making for patients with clinically significant PCa by improving diagnostic accuracy, particularly in minimizing intra- and inter-observer variability. Despite its high specificity, improvements in sensitivity and segmentation accuracy are needed, which could be achieved by using larger datasets and advanced deep learning techniques. Further multicenter validation is required for accelerated integration of this system into clinical practice.
About the Autors
Corresponding author: He Mingze – postgraduate student of the Institute for Urology and Reproductive Health, Sechenov University, Moscow, Russia; e-mail: hemingze97@gmail.com
Similar Articles
