MRI-based deep learning combined with radiomics for the preoperative prediction of lymphovascular invasion in patients with bladder cancer.

Abstract

Background: Lymphovascular invasion (LVI) signifies poor prognosis in bladder cancer, yet reliable preoperative prediction remains challenging, limiting personalized treatment planning.

Methods: In this multi-center retrospective study, 543 bladder cancer patients were enrolled. Of these, 473 patients from two centers were randomly split into training and internal test sets at an 8:2 ratio, while 70 patients from six additional centers constituted an independent external test set. After tumor and peritumoral segmentation, a hybrid model that integrates deep learning and radiomics features was developed. The hybrid model was compared against ten baseline models, including image-only and radiomics-only deep learning models, as well as radiomics-based machine learning approaches. Model performance was assessed using the area under the receiver operating characteristic curve (AUC) with five-fold cross-validation. Prognostic value was evaluated by Kaplan-Meier survival analysis with the log-rank test.

Results: The hybrid model achieved AUCs of 0.77 (internal test) and 0.75 (external test), surpassing all baseline models. The model-predicted LVI status significantly stratified overall survival in the training set (p < 0.001) and the internal test set (p = 0.003), with a non-significant trend observed in the external test set (p = 0.151).

Conclusion: The proposed non-invasive hybrid model can accurately predict LVI status preoperatively and demonstrates prognostic potential, thereby aiding risk stratification in bladder cancer.