Feature map quantification: An efficient approach for active trachoma image classification.

Area of Science:

• Computer Science
• Medical Imaging
• Artificial Intelligence

Background:

• Convolutional neural networks (CNNs) are effective for classifying active trachoma from inverted eyelid images.
• Deploying complex CNNs in medical centers is hindered by computational resource limitations.

Purpose of the Study:

• To propose a novel filter pruning framework (FSIM-SVD) to address computational constraints in CNNs for medical image analysis.
• To enable efficient deployment of deep learning models in resource-limited medical environments.

Main Methods:

• Developed a quantified feature map-based filter pruning framework (FSIM-SVD).
• Utilized feature similarity (FSIM) to quantify redundant feature maps.
• Employed singular value decomposition (SVD) to assess feature map contributions.
• Pruned less significant filters based on their impact on model performance.

Main Results:

• VGG16 achieved 86.9% accuracy for active trachoma classification, reducing FLOPs by 28.6% and parameters by 33.4%.
• ResNet110 achieved 94.31% accuracy on CIFAR10, with a 43.8% reduction in FLOPs and 43.1% reduction in parameters.
• The FSIM-SVD approach demonstrated higher pruning rates and improved classification performance compared to state-of-the-art methods.

Conclusions:

• The FSIM-SVD framework effectively reduces computational complexity of CNNs for medical image classification.
• This approach enhances the practicality of AI-driven diagnostic tools in resource-constrained healthcare settings.
• FSIM-SVD offers a promising solution for efficient and accurate deep learning model deployment in medical services.