A meta-fusion RCNN network for endoscopic visual bladder lesions intelligent detection

Area of Science:

• Medical Imaging
• Computer Vision
• Artificial Intelligence

Background:

• Endoscopic diagnosis of bladder lesions can be challenging due to limited sample sizes for training deep learning models.
• Accurate and efficient detection of bladder lesions is crucial for timely medical intervention.

Purpose of the Study:

• To develop and evaluate a novel deep learning-based visual object detection technology for assisting in the diagnosis of bladder lesions.
• To address the challenge of insufficient endoscopic lesion samples in deep neural network training.

Main Methods:

• A new object detection approach derived from cascade R-CNN, featuring a feature adaptive fusion model.
• Integration of task adaptation meta-learning to train feature fusion and network updates for adaptive classification and detection.
• Evaluation on standard datasets (Pascal VOC, Microsoft COCO) and a custom bladder lesion dataset.

Main Results:

• The proposed method demonstrated superior performance compared to the original cascade R-CNN on benchmark datasets.
• The model effectively adapted to limited endoscopic lesion samples, reducing overfitting.
• Successful application to a custom bladder lesion dataset confirmed its effectiveness in auxiliary detection for intelligent diagnosis.

Conclusions:

• The developed deep learning model significantly enhances visual object detection for bladder lesions in endoscopy.
• This technology shows promise for improving the accuracy and efficiency of intelligent bladder lesion diagnosis.