Multi-style spatial attention module for cortical cataract classification in AS-OCT image with supervised contrastive...

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Precise classification of cortical cataracts (CC) is crucial for timely intervention and surgical planning.
• Anterior segment optical coherence tomography (AS-OCT) shows promise for cataract diagnosis, but automatic classification of CC remains challenging due to complex opacity distributions.

Purpose of the Study:

• To explore CC opacity distribution characteristics as clinical priors to improve deep convolutional neural network (CNN) performance in CC classification.
• To enhance the representational capability and explainability of CNNs for AS-OCT-based CC classification.

Main Methods:

• Proposing a novel Multi-style Spatial Attention module (MSSA) for recalibrating CNN feature maps based on clinical contexts.
• MSSA integrates Group-wise Style Pooling (GSP), Local Transform (LT), and Style Feature Recalibration (SFR) for refined feature map recalibration.
• The MSSA module is designed for easy integration into existing CNN architectures with minimal computational overhead.

Main Results:

• Extensive experiments on CASIA2 AS-OCT and two public ophthalmic datasets demonstrated MSSA's superiority over state-of-the-art attention methods.
• Visualization analysis and ablation studies confirmed MSSA's enhanced explainability in the decision-making process.

Conclusions:

• The proposed MSSANet effectively utilizes CC opacity distribution characteristics to boost CNN representational power and explainability.
• This method shows significant potential for improving early clinical diagnosis of cortical cataracts.