Exploring transfer learning for gastrointestinal bleeding detection on small-size imbalanced endoscopy images.

Area of Science:

• Medical Imaging
• Computer Vision
• Machine Learning

Background:

• Convolutional Neural Networks (CNNs) excel at image representation but require extensive data.
• Traditional methods rely on hand-crafted features, which are less effective for complex tasks like medical image analysis.
• Small and imbalanced datasets pose significant challenges for training deep learning models.

Purpose of the Study:

• To explore transfer learning for gastrointestinal bleeding detection using limited endoscopy images.
• To demonstrate the efficiency of transferring CNN-learned representations to new tasks with small datasets.
• To improve the performance of CNNs on imbalanced medical image datasets.

Main Methods:

• Utilized a pre-trained Inception V3 model (trained on ImageNet) for feature extraction.
• Fine-tuned the pre-trained model using labeled endoscopy images, resuming training from learned weights.
• Implemented data augmentation and image resampling to increase dataset size and positive sample rate.

Main Results:

• Transfer learning significantly improved performance metrics including AUC, Precision, Recall, and Accuracy.
• The proposed transfer learning method outperformed both hand-crafted feature methods and training CNNs from scratch.
• Achieved superior results in detecting gastrointestinal bleeding on small, imbalanced datasets.

Conclusions:

• Transfer learning is a viable and effective strategy for medical image analysis tasks with limited data.
• Leveraging pre-trained models and augmentation techniques enhances CNN performance in challenging scenarios.
• This approach offers a practical solution for improving diagnostic accuracy in gastrointestinal endoscopy.