A Multi-Label Detection Deep Learning Model with Attention-Guided Image Enhancement for Retinal Images

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Multi-disease fundus image classification faces challenges with limited data, uneven distributions, and low accuracy.
• Deep learning models typically require large datasets, which are often unavailable for specific medical tasks.

Purpose of the Study:

• To develop an ensemble model for multi-disease fundus image classification that overcomes data scarcity.
• To improve the accuracy and robustness of fundus image classification using deep learning.

Main Methods:

• Proposed an ensemble model combining VGG19 and ResNet50 classification networks.
• Integrated Gradient-weighted Class Activation Mapping (Grad-CAM) as a data augmentation module to generate activation maps.
• Utilized both original and augmented data, along with model fine-tuning and transfer learning.

Main Results:

• Achieved an accuracy of 97%, precision of 92%, and recall of 81% on the Retinal Fundus Multi-Disease Image Dataset (RFMiD).
• The Grad-CAM module guided the model to learn lesion feature differences, enhancing robustness.
• Activation maps provided interpretability, highlighting areas of interest for classification.

Conclusions:

• The proposed Grad-CAM-based ensemble model effectively addresses data limitations in multi-disease fundus image classification.
• The method demonstrates superior performance compared to other approaches, offering a robust and interpretable solution.
• This technique has the potential to advance automated diagnosis in retinal imaging.