Multi-DECT image-based intra- and peritumoral interpretable radiomics and artificial intelligence models for preoperative prediction of pathological grading of non-muscle-invasive bladder cancer.

Authors

Abstract

This study aimed to evaluate the predictive model based on multi-dual-energy computed tomography (DECT) imaging parameters and model integrating DECT parameters with intra- and peritumoral- radiomics and deep learning model using a 2.5D Vision Transformer (ViT) for the preoperative prediction of pathological grading in non-muscle-invasive bladder cancer (NMIBC). This retrospective study included 193 patients with NMIBC (130 with high-grade and 63 with low-grade by pathological biopsy), who were randomly divided into training and test cohorts in an 8:2 ratio. The virtual monoenergetic images (VMIs) at 40 keV, 70 keV, and 100 keV, and iodine material decomposition (IMD) images were used for analyses. Univariate and multivariate logistic regression analyses were conducted to identify independent predictors of pathological grade in NMIBC, formed a DECT model. Radiomics and 2.5D ViT features were concurrently extracted from the intratumoral and peritumoral regions on these images. Multilayer perceptron (MLP) classifiers were utilized to develop models for intratumoral, peritumoral, and combined intratumoral and peritumoral (IntraPeri) radiomics and ViT features. SHapley Additive exPlanations (SHAP) were employed to visualize the decision-making process of the optimal model. Additionally, the independent predictors from DECT were integrated with the optimal model to develop a predictive nomogram. The diagnostic performance of the models was evaluated using receiver operating characteristic (ROC) curves, while decision curve analysis (DCA) was applied to assess their potential clinical utility. Normalized iodine concentration (NIC) was identified as an independent predictor of pathological grade in NMIBC with an area under the curve (AUC) of 0.698 (95% CI: 0.528-0.868) in the test cohort, while the combined intra- and peritumoral model achieved a higher AUC of 0.864 (95% CI: 0.739-0.989). The nomogram incorporating DECT independent predictors and combined model yielded a numerically higher AUC of 0.891 (95% CI: 0.786-0.996), suggesting a trend toward improved predictive performance. However, the pairwise DeLong test between the DECT model and the nomogram did not reach statistical significance (p = 0.055) in the test cohort. Calibration curves and DCA indicated a good model fit and net benefit in this cohort. An interpretable model combining DECT multiparametric images and their intra- and peritumoral radiomics may aid in predicting the pathological grade of NMIBC. These preliminary findings, derived from a single-center retrospective dataset, require confirmation in larger, externally validated cohorts.

Tags

Topics