Abstract
The proposed DL model based on CT images exhibited relatively good prediction ability of muscle-invasive status of BCa preoperatively, which may improve individual treatment of BCa.
Introduction
Determining the muscular invasiveness of bladder cancer (BCa) is a critical diagnostic step that dictates surgical strategy. Non-Muscle-Invasive Bladder Cancer (NMIBC) typically allows for local resection, whereas Muscle-Invasive Bladder Cancer (MIBC) often necessitates radical cystectomy and chemotherapy. Currently, the gold standard—Transurethral Resection of Bladder Tumor (TURBT)—is operator-dependent and carries a misdiagnosis rate of up to 25% for MIBC. While CT imaging is standard, human radiologists struggle to distinguish subtle bladder wall layers. This study introduces a deep learning approach to extract discriminative features from preoperative CT scans to predict invasiveness non-invasively.
Methodology
The researchers developed the Filter-guided Pyramid Network (FGP-Net), a 3D Convolutional Neural Network designed to capture both global tumor features and local margin details.
Key technical innovations include:
- Two-Channel Input: The model processes both the original tumor region and a masked version, focusing attention on the tumor-bladder wall interface.
- FGP-Net Architecture: Designed to integrate multi-scale features, ensuring the model considers the entire tumor mass alongside fine-grained edge characteristics.
- Focal Loss: This loss function was utilized to address class imbalance, preventing bias toward the more frequent NMIBC cases.
- Explainability: t-SNE visualizations and feature maps were used to confirm that the model's decisions were based on clinically relevant areas, such as the perivesical fat and internal tumor regions.
Patient Cohorts
The study utilized retrospective data from 441 patients across two centers to ensure robust development and validation.
- Center 1 (Development & Internal Validation): 366 patients were split into development (n=183), tuning (n=110), and internal validation (n=73) sets.
- Center 2 (External Validation): 75 patients were used to test the model's generalizability across different scanners and protocols.
Results and Analysis
The FGP-Net model demonstrated superior performance compared to traditional radiological assessment:
- Predictive Accuracy: The model achieved an AUC of 0.861 in internal validation and 0.791 in external validation.
- Human Comparison: The AI significantly outperformed two experienced radiologists, who achieved AUCs of 0.685 and 0.585 in the same internal cohort. Radiologists tended to "over-stage" tumors, whereas the DL model provided a more balanced assessment.
- Clinical Utility: Decision Curve Analysis (DCA) confirmed that the DL model offers a higher net benefit for clinical decision-making than current standard practices.
- Size Correlation: Analysis showed the model was particularly effective for tumors larger than 4cm, while smaller lesions (<1cm) remained a challenge for both AI and humans.
Conclusion
This study demonstrates that deep learning can successfully identify "hidden" features in CT scans to predict bladder cancer invasiveness. While retrospective, the results suggest that FGP-Net could serve as a valuable tool for preoperative staging, potentially reducing the need for invasive biopsies or preventing under-treatment. Future work will focus on prospective trials and integrating multi-modal data, such as clinical biomarkers, to further refine diagnostic precision.
Summary updated