David Harris-Birtill

PhD, MPH

University of St Andrews

Senior Lecturer

PLCO (Learn more about this study)

PLCOI-1802

Jan 27, 2025

Deep Learning for Bladder Cancer Detection and Segmentation

Bladder cancer poses a clinical challenge due to its high recurrence rates and diverse presentations. Accurate early detection is critical to improving patient outcomes. Deep learning–based methods, convolutional neural networks, emerging transformer architectures or hybrid architectures, have shown promising results in detecting and segmenting bladder tumors. However, questions remain about how to handle variability in tumor appearance and enhance interpretability for clinicians.

Building on recent progress, this project aims to advance current diagnostic capabilities for bladder cancer by developing a pipeline for automated bladder cancer detection and segmentation. It will begin with a review of successful approaches. This will guide the creation of a baseline detection model. Then it will build an advanced system that integrates transformer-based methods and data augmentation techniques to adapt to heterogeneous clinical conditions. The project aims to create a model with both high sensitivity and specificity.

The PLCO bladder cancer dataset would be instrumental for this project. Its comprehensive collection of histopathology slides and patient data enables to correlate imaging-based predictions with definitive tissue-level findings and longitudinal outcomes. This enables cross-validation of different detection algorithms. It also supports the development of an explainable model that physicians can trust.

Thank you very much for considering this application.

Overall Aim: Advance current diagnostic capabilities for bladder cancer.

1. Review of existing deep learning approaches for bladder cancer detection
- Identify current gaps and limitations

2. Baseline Model Development
- Implement minimal viable detection model (e.g., CNN-based classifier) to identify bladder tumors
- Establish performance baselines for sensitivity, specificity, and computational efficiency

3. Advanced Pipeline for Detection and Segmentation
- Develop segmentation model (e.g., U-Net, Vision Transformer, Hybrid Architecture) for bladder tumor delineation
- Integrate interpretability tools (e.g., Saliency maps, Grad-CAM) for clinically meaningful visual explanations
- Tune hyperparameters (learning rate, batch size, depth, etc.)

4. Comparative Analysis
- Benchmark against other existing models and methods in bladder cancer detection and segmentation