Prostate Cancer Classification and Interpretation With Multiparametric Magnetic Resonance Imaging and Gleason Grade Score Using DarkNet53 Model

• Published in: The Prostate 2025-02, Vol.85 (3), p.294-307
• Main Authors: Gladis Pushparathi, Vasantha Pragasam, Justin Xavier, Dhas, Chitra, Pandian, Kannan, Gopalraj
• Format: Article
• Language: English
• Subjects: Algorithms; Automation; Classification; Classification systems; DarkNet53 model; Grad‐CAM; Humans; Image Interpretation, Computer-Assisted - methods; Image processing; Magnetic resonance imaging; Male; Multiparametric Magnetic Resonance Imaging - methods; Neoplasm Grading - methods; Noise reduction; Prostate cancer; Prostatic Neoplasms - classification; Prostatic Neoplasms - diagnostic imaging; Prostatic Neoplasms - pathology; weighted mean of vectors optimization
• ISSN: 0270-4137; 1097-0045; 1097-0045
• DOI: 10.1002/pros.24827

ABSTRACT Background Prostate Cancer (PCa) increases the mortality rate of males worldwide and is caused by genetics, lifestyle, and age reasons. The existing automated PCa classification systems face difficulties with overfitting issues, and non‐generalizability, leading to poor classification performance. Objective On this account, this study proposes an automated classification of PCa from MRI images using a hybrid weighted mean of vectors‐optimized DarkNet53 classifier model. Methodology The proposed method suggests nonlocal mean filtering for noise reduction, N4ITK bias field correction to enhance image quality, and active contour‐based segmentation for accurately identifying the disease region. The feature extraction utilizes the gray level run length matrix and shape features for effective feature extraction. A weighted mean of vectors optimization is used to optimize the feature selection process by hybridizing it with the DarkNet53 model for classification. Finally, the interpretation of achieving the classification has been demonstrated using the explainable AI Grad‐CAM model. Results After comparing the proposed work with various state‐of‐the‐art algorithms, the proposed model achieves 99.31% accuracy, 98.24% sensitivity, and 98.46% specificity, respectively, highlighting the model's accomplishment using the DarkNet53 classifier.