Multimodal deep learning framework integrating multiphase CT and histopathological whole slide imaging for predicting recurrence in ccRCC.
Authors
Affiliations (8)
Affiliations (8)
- Department of Radiology, Jiangmen Central Hospital, 23 Beijie Haibang Street, Jiangmen, People's Republic of China.
- School of Electronic Information and Automation, Guilin University of Aerospace Technology, 2 Jinji Road, Guilin, People's Republic of China.
- Jiangmen Key Laboratory of Artificial Intelligence in Medical Image Computation and Application, 23 Beijie Haibang Street, Jiangmen, People's Republic of China.
- Department of Pathology, Jiangmen Central Hospital, 23 Beijie Haibang Street, Jiangmen, People's Republic of China.
- Guangdong Medical University, 2 Wenming East Road, Zhanjiang, People's Republic of China.
- Department of Radiology, Jiangmen Central Hospital, 23 Beijie Haibang Street, Jiangmen, People's Republic of China. [email protected].
- Jiangmen Key Laboratory of Artificial Intelligence in Medical Image Computation and Application, 23 Beijie Haibang Street, Jiangmen, People's Republic of China. [email protected].
- Guangdong Medical University, 2 Wenming East Road, Zhanjiang, People's Republic of China. [email protected].
Abstract
ccRCC is an aggressive, heterogeneous tumor with a poor prognosis. Prognostic assessments need multi-modal data. Radiological images have limits, while pathological images offer micro-level details. Integrating these for ccRCC outcome prediction is important. Our study aimed to develop and validate a DL fusion model using multiphase CT images and WSI for postoperative risk stratification in ccRCC patients. This retrospective study included 274 ccRCC patients who underwent multiphase CT scans (Jan 2008-Mar 2021), with diagnoses confirmed by histopathology post-surgery. The patient cohort was divided into a training cohort of 164 patients for model development and a test cohort of 110 patients for model validation. The primary outcome was local recurrence or metastasis versus non-recurrence (NR) with a minimum follow-up of 3 years. DL models based on multiphase CT images and histopathological WSIs were developed and validated. Performance comparisons among models were made through accuracy (ACC) and receiver operating characteristic (ROC) curve analyses, with integrated discrimination improvement (IDI) analysis and the DeLong test assessing diagnostic performance. Decision curve analysis (DCA) evaluated clinical utility, and Kaplan-Meier analysis assessed variable-survival correlations. The CT and Pathology Mutual Guidance Fusion Diagnostic Network (CPNet) exhibited superior performance in predicting postoperative disease-free survival (DFS) in ccRCC patients. Among the models, the PCP-Pathology Fuse model achieved the highest AUC of 0.8363 and accuracy of 75.45%, outperforming the CMP-Pathology Fuse (AUC 0.7965, ACC 69.09%) and NP-Pathology Fuse (AUC 0.798, ACC 69.09%) models. Its performance was comparable to the Three-phase-Pathology Fuse model (AUC 0.8341, ACC 70.00%, P > 0.05). IDI and DCA confirmed significant net benefits (0.01-0.95) for the PCP-based model. The PCP-based CPNet model shows promise for predicting postoperative DFS in ccRCC patients, with performance comparable to three-phase CT-pathology models. It may serve as a potential bioimaging prognostic marker, pending external validation to support clinical integration.