Bladder cancer is a serious public health concern due to the high rate of recurrence and substantial variation in patient outcomes. It is important to accurately predict how a patient will respond to chemotherapy, as it allows clinicians to guide treatment plans personalisation and avoid unnecessary toxicity. [3]. Therefore, there is an increasing demand for non-invasive methods to predict treatment outcomes, which could improve the quality of life of bladder cancer patients and optimize therapeutic decision-making.
ML methods combined with MRIT have emerged as a promising tool in oncology for non-invasive prediction of treatment response [5]. Radioactive allows for the extraction of quantitative features from medical imaging modalities that capture tumor heterogeneity and other robust features that are invisible to the human eye [10]. By applying advanced ML algorithms, researchers aim to leverage these features to create models that can predict treatment effects before chemotherapy [26]. These models can benefit bladder cancer patients and provide a path to individualized treatment and potentially better survival [5]. New contributions in our research include the incorporation of multiparametric MRI sequences, the application of RAD score feature engineering, and comparative analysis between multiple ML classifiers.
In a previous study of MIBC patients, Cha et al. [19] We developed a CT scan-based radioactive model to predict NAC responses using 91 radiomics features. Testing the optimal model achieved an AUC value of 0.77 when predicting post-treatment CR. In subsequent studies by the same group, Wu et al. [17] We utilized transfer learning to train the model using CT scans from 123 MIBC patients undergoing NAC, resulting in a test AUC of 0.79. In another study by Choi et al. [18]a radioactive signature derived from CT scans was developed to predict pathological CR using an RF classifier, achieving excellent performance on the training data set (AUC, 0.85) and moderate performance on the validation (AUC, 0.75) set. Comparative analysis of these previous radiomics studies and current studies is described in Supplementary Table 2 (Supplementary File 2). This table positions MRI-derived AUCs within the context of previous CT-based studies, highlighting the value-added of MRI sequences for improved differential performance and soft tissue characterization. Additionally, Supplementary Table 3 (Supplementary File 2) provides a comparison of the architecture of rad scores and classical ML pipelines and commonly used note-based pipelines. This table shows that the methodological novelty of our study lies in the systematic multiclassification assessment with prospective acquisition of MRI data, standardized IBSI-compliant feature extraction (MIRP), transparent rad-score equations for clinical interpretability, and robust cross-validation.
In this study, we developed an ML-based radiomic model using four MR Image series (CE-T1WI, T2WI, DWI, and ADC MAP) to predict clinical CR in NACs in MIBC patients. The results demonstrated that different imaging series and clinical features provide varying degrees of predictive power.
Among the various image series, CE-T1WI exhibited excellent predictive performance, with classifiers such as SVM and KNN achieving the highest AUC-ROC values (0.88 and 0.87, respectively). This suggests that this sequence captures important tumor properties such as tumor shape, margins, elongation, heterogeneity, and vascular permeability, which are important for assessing treatment response. In contrast, T2WI shows low predictive power, and while this sequence provides valuable information, additional optimization or functional engineering may be required to enhance the utility. It also highlights the need for a multimodal approach, as demonstrated in large-scale studies that integrate different imaging techniques to enhance predictive power. [8].
Using clinical features for response prediction resulted in excellent model performance as RF and SVM classifiers achieved the highest prediction accuracy. Baseline clinical features including tumor volume, T-stage, and nodular status remain fundamental contributors to the predictive model. These findings are consistent with previous radioactive studies that emphasize the inclusion of clinical data in predictive models. [27, 28]. With regard to statistical considerations, a relatively large dataset is required to integrate clinical and radioactive functions into a single model. Given the limited size of the main dataset, this study developed a clear model of various clinical and imaging functions.
Among the various classifiers used in our study, the SVM algorithm consistently outperformed the others, achieving an AUC of 0.88. This is consistent with existing research supporting the utility of SVM in radiomics, highlighting its ability to capture complex patterns in imaging data [29,30,31,32]. For example, studies by Du et al. [33] SVM demonstrated that when applied to radioactive features extracted from pretreatment MRI, it could significantly predict the treatment response of brain metastasis patients to stereotactic radios, further highlighting the robustness of this classifier in oncology. Furthermore, Ji et al. [34] The SVM classifier was found to effectively predict lymph node metastasis in patients with vesicourothelial carcinoma treated with radical cystectomy and enhance the results of this analysis. As with our findings, Cai et al. [8] We highlighted the promising role of MRI-based radiation using an SVM classifier in predicting response to NAC in cervical cancer patients. Their results demonstrated that MRI-derived radiation can identify key features associated with tumor heterogeneity and vascular permeability, and achieve an AUC of 0.86 for predicting treatment response. Similarly, the AUC of a study of 0.88 CE-T1WI-based radiation with SVM supports the high discriminatory power of MRI sequences to predict CR.
In our study, the LGBM classifier demonstrated the lowest predictive performance for assessing chemotherapy response across T2WI, DWI, and clinical features. Several factors may explain this outcome. LGBM is known for its efficiency over large datasets due to gradient-based one-sided sampling and exclusive functional bundles. This is very effective when there is rich, high-dimensional data. However, for small or unbalanced datasets, as often encountered in Radiomics studies, LGBM algorithms may not effectively capture subtle patterns, potentially leading to performance. Comparing this finding with previous studies, several studies report favorable performance of LGBMs to predict early complications after radical gastrectomy and breast cancer classification, especially when a broad data set is available. [35, 36]. This contrast suggests that LGBM applicability may be context-dependent, and alternative models such as SVM and RF may be more suited to smaller and higher dimensional radiation data. Future studies should consider further investigation of dataset size and balance requirements for optimal use of LGBM in predictive modeling.
Clinical implications
Our findings suggest that MRI-based radioactive models combined with RAD scores and ML classifiers may serve as valuable decision support tools in the management of MIBC patients. By enabling non-invasive and pre-treatment prediction of response to NAC at the patient level, this approach helps clinicians identify candidates who are likely to benefit from NAC. Transparency in the rad score further promotes interpretability and allows physicians to understand which imaging features contribute to individual predictions. Although additional validation is required in a large multicenter cohort, the model may support personalized treatment plans and complement interdisciplinary decision-making.
limit
Despite these promising results, this study has some limitations. First, sample sizes are sufficient for initial model development, but may limit the generalizability of the findings to a larger population. Second, external validation of independent cohorts is required to ensure the reproducibility and clinical applicability of the developed models. Third, class weighting, stratified cross-validation, and Lasso feature selection have helped to reduce overfitting and imbalances, but these techniques may not completely eliminate bias in small datasets. Fourth, some models need to consider a relatively wide range of 95% CI. The narrow CI suggests that the AUC values are probably accurate, while the wide CI indicates that the AUC values are less reliable [25].
Future direction
Future research could benefit from the integration of advanced deep learning methods and the use of longitudinal imaging to capture temporal dynamics across the NAC. Furthermore, establishing multicenter collaboration allows access to larger datasets, supports external validation, and enhances the robustness and clinical applicability of ML-based radiomics models.
