DeepBCTPred: deep learning-based prediction of bladder cancer tissues from endoscopic images

Area of Science:

• Oncology
• Artificial Intelligence
• Medical Imaging

Background:

• Bladder cancer is a prevalent malignancy with diagnostic challenges due to subjective interpretation of conventional methods.
• Accurate tissue classification is crucial for early diagnosis, patient survival, and quality of life.
• Existing diagnostic techniques are limited by subjectivity and potential for human error.

Purpose of the Study:

• To develop and evaluate DeepBCTPred, a novel deep learning framework for accurate bladder cancer tissue classification.
• To integrate handcrafted and learned features using a dual-branch architecture for enhanced diagnostic performance.
• To improve upon the accuracy and reliability of current bladder cancer diagnostic methods.

Main Methods:

• A dual-branch deep learning architecture combining MobileNetV3 and a Feedforward Neural Network was developed.
• Recursive Feature Elimination (RFE) was employed for effective feature selection.
• A genetic algorithm-based image generation pipeline was utilized for optimal data selection and augmentation.

Main Results:

• DeepBCTPred achieved high performance metrics: 98.74% recall, 99.45% specificity, and 97.96% F1-score on the test dataset.
• The framework significantly outperformed existing state-of-the-art methods, with recall improvements of 2%-15%.
• Substantial gains were observed in F1-score (1.3%-13.1%) and Matthews Correlation Coefficient (MCC) (1.5%-16%).

Conclusions:

• DeepBCTPred demonstrates superior performance in bladder cancer tissue classification, offering a significant advancement over current methods.
• The framework shows strong potential for clinical implementation in bladder cancer diagnosis.
• The developed deep learning approach may be extensible to other cancer types, contributing to precision medicine applications.