What question did this study set out to answer?

The aim is to develop a radiomics model that predicts bladder cancer grade from white-light cystoscopy images.

April 17, 2026Open Access

A Novel Radiomics-based Interpretable Model for Bladder Cancer Grade Prediction Using White-Light Cystoscopy Images

Puntos clave

The aim is to develop a radiomics model that predicts bladder cancer grade from white-light cystoscopy images.
Developed a radiomics model using multichannel images from white-light cystoscopy.
Segmented 2624 tumour regions from images of 423 patients for training and validation.
Applied machine-learning classifiers with feature selection using coefficient thresholding and LASSO.
Assessed model performance through discrimination, calibration, and decision curve analysis (DCA).
Evaluated interpretability using SHAP analysis and feature visualisation.
The support vector machine model achieved an AUC of 0.87 for internal validation and 0.79 for external validation.
Distinct radiomic patterns identified by SHAP analysis differentiated low-grade from high-grade tumours.
Calibration acceptable at net benefit thresholds ≥ 0.30, but external validation highlighted limitations.

Resumen

A novel radiomics model using cystoscopic images reliably predicts bladder-cancer grade and offers interpretable visual patterns, potentially reducing invasive procedures and accelerating diagnosis in clinical practice. White-light cystoscopy (WLC) is the standard diagnostic modality for bladder cancer, but preoperative grading remains inaccurate. We developed a multichannel radiomics model to predict tumour grade (low-grade LG vs high-grade HG) from WLC and to identify imaging biomarkers. WLC images were retrospectively collected from 423 patients across two centres. A total of 2624 tumour regions were segmented for training, with 584 and 358 regions for internal and external validation, respectively. Radiomic features were extracted from the greyscale and red–green–blue channels. Feature selection was performed using coefficient thresholding and the least absolute shrinkage and selection operator. Five machine-learning classifiers were trained. Model performance was assessed using discrimination, calibration, and decision curve analysis (DCA). Interpretability was assessed using SHapley Additive exPlanations (SHAP) and feature visualisation. The support vector machine model achieved robust performance, with an area under the receiver operating characteristic curve of 0.87 (95% confidence interval CI = 0.84–0.89) for internal validation and 0.79 (95% CI = 0.73–0.85) for external validation. SHAP analysis revealed distinct radiomic patterns differentiating LG from HG tumours. Limitations include retrospective design, manual segmentation, and a small, imbalanced external set, so validation reflects preliminary transportability rather than robustness or generalisability. Although calibration was acceptable and net benefit appeared at thresholds ≥ 0.30, external data constraints warrant caution. The proposed multichannel radiomics model supports grade prediction from WLC images and identifies a green channel. This approach provides a basis for developing real-time, filter-based tools for intraoperative risk stratification.

Leer artículo completoexternamente

Me gusta

Guardar

Ver artículo completo