A comprehensive framework for multi-class prostate cancer classification using biparametric MRI: a multi-center retrospective study.
Authors
Affiliations (10)
Affiliations (10)
- Department of Diagnostic Imaging, Nanfang Hospital, Southern Medical University, Guangzhou, China.
- Department of Medical Imaging, Zhongshan Hospital of Traditional Chinese Medicine, Guangdong, China.
- Department of Urology, Nanfang Hospital, Southern Medical University, Guangzhou, China.
- Department of Urology, Guangdong Provincial People's Hospital, Guangdong Academy of Medical Sciences, Southern Medical University, Guangzhou, China.
- Medical Imaging Department of Ganzhou People's Hospital, Ganzhou, China.
- School of Automation, Guangdong University of Technology, Guangzhou, China.
- Guangdong Provincial Key Laboratory of Intelligent Decision and Cooperative Control, Guangzhou, China.
- Guangdong-Hong Kong Joint Laboratory for Intelligent Decision and Cooperative Control, Guangzhou, China.
- Department of Plastic Surgery, The First Affiliated Hospital of Ningbo University, Ningbo, China. [email protected].
- Department of Urology, Guangdong Provincial People's Hospital, Guangdong Academy of Medical Sciences, Southern Medical University, Guangzhou, China. [email protected].
Abstract
Accurate pre-biopsy or preoperative diagnosis and risk stratification of prostate cancer remain challenging. This study aimed to develop an automated framework for prostate and suspicious lesion segmentation, feature extraction, and multiclass classification of prostate cancer. This study enrolled 2543 patients with suspected clinically significant prostate cancer (csPCa). An nnUNet was used for prostate segmentation, followed by suspected lesion identification and feature extraction. Unimodal and multimodal deep learning models were trained to classify patients into non-csPCa (benign/ISUP 1), mid-risk csPCa (ISUP 2-3), and high-risk csPCa (ISUP 4-5). Subsequently, we extracted prostate radiomics, lesion radiomics, and learned features from the unimodal DL model. Radiomics features, learned features from the unimodal deep learning model, and clinical variables were integrated to develop machine-learning classifiers. The best-performing model was compared with conventional clinical variables and evaluated across subgroups defined by PSA levels and PI-RADS 4. The nnUNet achieved scores of 0.971 in internal validation and 0.934 in external testing. The optimal machine learning model, which integrated all features, achieved a macro-AUC of 0.87 [0.84-0.90] across both internal and external evaluations. Further comparisons indicate it outperforms traditional metrics such as total PSA, PSAD, and PI-RADS. Additionally, it exhibits strong robustness in most subgroup analyses, consistently surpassing conventional variables. We developed an automated bpMRI-based framework that segments the prostate and suspicious lesions and enables multiclass risk stratification of prostate cancer by integrating imaging and clinical features. The proposed approach demonstrated strong performance and robustness, outperforming conventional clinical variables. Question accurate pre-biopsy stratification of prostate cancer is essential for clinical decision-making. Current studies predominantly focus on binary classification of csPCa, and multi-class classification remains challenging. Findings the comprehensive framework achieved an area under the curve of 0.87 in both internal and external tests, exceeding clinical variables and demonstrating robustness. Critical relevance statement this study presents a robust multi-class classification framework based on biparametric MRI and machine learning to improve pre-biopsy prostate cancer stratification, outperform conventional metrics, and provide a valuable tool for risk stratification and clinical decision-making.