•  
  •  
 

Corresponding Author

V. Gokula Krishnan

Subject Area

Biomedical Engineering

Article Type

Original Study

Abstract

Bladder cancer diagnosis demands the integration of cystoscopy, histopathology, and biomarker testing, yet the majority of AI systems continue to be modality, specific and sensitive to acquisition variations. TRIDENT, Bladder is a resolution, aware multimodal framework that we have developed to address this issue. The method combines cystoscopic lesion segmentation, whole, slide image (WSI) molecular subtype prediction, and urine miRNAclinical data through reliability, weighted fusion and a utility, aware decision policy. The endoscopic branch features a quality, gated super, resolution component, resulting in mDice 87.5, sensitivity 92.1%, precision 91.4%, boundary, F1 85.3%, and ECE 0.023 at 32 FPS. mDice on low, quality frames increases from 65.1 to 72.4 while ECE drops from 0.081 to 0.036. For histopathology, attention, based MIL with CORAL alignment produces AUROC 0.90 (TCGA), 0.85 (STPH), and 0.74 (GD2H), thus boosting cross, site generalization by up to +0.10. The liquid biopsy/clinical model obtains AUROC 0.86, F1 0.80, Brier 0.128, and ECE 0.021 with millisecond, level CPU inference. Reliability, weighted fusion reaches AUROC 0.94, PR, AUC 0.92, ECE 0.019, and FN@95%Spec 0.08, surpassing single, modality and equal, weight baselines. Cost, derived thresholds (t = 0.35, t = 0.70) are used to maximize the expected clinical utility. The findings indicate enhanced robustness, calibration, and decision value over a range of clinical scenarios

Keywords

Bladder Cancer; Whole-slide image; Super-resolution; Liquid-biopsy; Domain alignment; Calibrated probabilities

Creative Commons License

Creative Commons Attribution 4.0 License
This work is licensed under a Creative Commons Attribution 4.0 License.

Download

Share

COinS