Tumor detection under cystoscopy with transformer-augmented deep learning algorithm.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Accurate bladder tumor detection is crucial for effective bladder cancer treatment and reducing recurrence.
• Standard white-light cystoscopy (WLC) can be enhanced by advanced deep learning algorithms for improved noninvasive diagnosis.

Purpose of the Study:

• To develop a cost-effective, transformer-augmented deep learning algorithm for precise bladder tumor detection in WLC.
• To evaluate the performance of this novel algorithm on archived patient data.

Main Methods:

• Developed 'CystoNet-T', a deep learning model featuring a transformer-augmented pyramidal Convolutional Neural Network (CNN) architecture.
• Integrated self-attention mechanisms via transformer encoder modules into the Feature Pyramid Network (FPN) for global feature aggregation.
• Trained and tested the model on WLC frames from patient cystoscopy videos, utilizing a dataset of 510 frames for training and 101 for testing.

Main Results:

• CystoNet-T achieved 96.4 F1 score and 91.4 Average Precision (AP) on the test set.
• Outperformed benchmark models Faster R-CNN and YOLO by 7.3 points (F1) and 3.8 points (AP).
• Demonstrated superior ability in highlighting foreground tumor information for accurate localization and minimizing false positives.

Conclusions:

• A novel deep learning algorithm, CystoNet-T, has been developed for accurate bladder tumor detection in WLC.
• The transformer-augmented AI framework shows significant promise for improving clinical decision-making in bladder cancer diagnosis and therapy.