An interpretable machine learning model based on CT imaging for predicting lymphovascular invasion and survival in bladder urothelial carcinoma: a multicenter study.
Authors
Affiliations (8)
Affiliations (8)
- Department of Urology, The First Affiliated Hospital of Chongqing Medical University, Chongqing, China.
- Department of Urology, Yibin Second People's Hospital, Yibin, China.
- Department of Urology, Yongchuan Hospital of Chongqing Medical University, Chongqing, China.
- Department of Radiology, The First Affiliated Hospital of Chongqing Medical University, Chongqing, China.
- Department of Urology, The Second Affiliated Hospital of Chongqing Medical University, Chongqing, China.
- Center for Reproductive Medicine, Women and Children's Hospital of Chongqing Medical University, Chongqing, China. [email protected].
- Chongqing Clinical Research Center for Reproductive Medicine, Chongqing Health Center for Women and Children, Chongqing, China. [email protected].
- Department of Urology, The First Affiliated Hospital of Chongqing Medical University, Chongqing, China. [email protected].
Abstract
Lymphovascular invasion (LVI) is a critical prognostic factor in bladder cancer, affecting recurrence, survival, and overall prognosis. Traditional methods for diagnosing LVI, such as immunohistochemical staining, are costly and time-consuming, making non-invasive alternatives like radiomics-based models valuable. This study aimed to construct an interpretable machine learning model to predict LVI status and survival outcomes in patients with bladder urothelial carcinoma using preoperative CT images. This study retrospectively enrolled patients with urothelial carcinoma who underwent radical cystectomy and preoperative contrast-enhanced CT from three medicine centers. Tumor regions were manually segmented, and radiomics features were extracted and selected through reproducibility testing, correlation analysis, and LASSO. Based on the selected radiomics features, machine learning classifiers, including SVM, were trained using five-fold cross-validation. A combined model was then constructed by integrating the radiomics signature with clinical risk factors. Model performance was evaluated by AUC, ACC, sensitivity, specificity, and survival analysis. The SVM model showed high performance, with an AUC of 0.944 in the training set and 0.872 in the testing set. The combined model integrating clinical factor performed better, achieving an AUC of 0.952 in the training set and 0.901 in the testing set. The model's interpretability was enhanced using SHAP analysis, identifying key radiomics features associated with LVI, such as tumor shape and texture. Survival analysis indicated that patients predicted to be LVI-negative had significantly better disease-free survival compared to patients predicted to be LVI-positive. This multicenter study demonstrates that the interpretable machine learning model based on preoperative CT images can effectively predict LVI status and survival outcomes in bladder urothelial carcinoma. This study was retrospectively registered by Ethics Committee of the First Affiliated Hospital of Chongqing Medical University (Approval No. K2024-187-01) on April 12, 2024, and informed consent was waived.