Combining Gabor-local and contextual-global deep features for cholesteatoma classification

Area of Science:

• Medical Imaging
• Computer Vision
• Artificial Intelligence

Background:

• Classifying cholesteatoma images requires effective extraction of both local and global features.
• Existing methods may struggle to integrate these features optimally.

Purpose of the Study:

• To develop a network that enhances the extraction and integration of local and global features for improved cholesteatoma image classification.
• To address the limitations of current feature extraction techniques in medical image analysis.

Main Methods:

• Introduction of a Gabor-local and contextual-global network.
• Utilizing Gabor filtering with ResNet for enhanced local feature extraction at various scales and orientations.
• Employing an attention mechanism with Generalized-Mean (GeM) pooling for robust global feature description.
• Implementing a dynamic feature fusion strategy based on computed importance scores.

Main Results:

• The proposed network demonstrates superior performance in cholesteatoma image classification.
• Effective integration of local and global features leads to enhanced classification outcomes.
• The dynamic fusion strategy optimizes the contribution of different feature types.

Conclusions:

• The Gabor-local and contextual-global network successfully addresses the challenge of integrating local and global features.
• This approach offers a more robust and accurate method for cholesteatoma image classification.
• The dynamic feature fusion strategy is key to the model's enhanced performance.