Retinopathy grading with deep learning and wavelet hyper-analytic activations.

Area of Science:

• Medical Imaging
• Computer Vision
• Artificial Intelligence

Background:

• Diabetic Retinopathy (DR) grading is crucial for patient care.
• Wavelet-based methods and Deep Learning, particularly Convolutional Neural Networks (CNNs), have advanced DR classification.
• Conventional CNN activation functions can discard valuable negative feature map information.

Purpose of the Study:

• To enhance feature representation in CNNs for improved DR grading.
• To introduce a novel Hyper-Analytic Wavelet (HW) phase activation function for wavelet sub-bands.
• To evaluate the effectiveness of the proposed activation function integrated with Multi-Resolution Analysis (MRA) and CNN models.

Main Methods:

• Integration of Multi-Resolution Analysis (MRA) with CNN frameworks.
• Formulation of a novel Hyper-Analytic Wavelet (HW) phase activation function that preserves negative coefficients.
• Application of the HW activation function to wavelet sub-bands within CNN architectures, including modified Alex Net for DR.
• Utilizing spatial-Wavelet quilts for feature enhancement.

Main Results:

• The proposed HW activation function effectively transforms negative coefficients, preserving significant edge features.
• CNN models incorporating spatial-Wavelet quilts and HW activations demonstrated superior performance.
• Modified Alex Net for DR achieved an 11% accuracy improvement (from 87% to 98%) with spatial-Wavelet quilts.
• The highest accuracy of 98% and sensitivity of 99% were attained with the Modified Alex Net for DR.

Conclusions:

• The integration of spatial-Wavelet quilts and hyper-analytic activations enhances the generalization ability of CNN models for DR grading.
• The novel HW activation function successfully preserves negative edge features, leading to improved classification accuracy.
• Visualization techniques confirmed better learning of feature maps from wavelet sub-bands.