Research on classification method of high myopic maculopathy based on retinal fundus images and optimized ALFA-Mix...

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• High myopic maculopathy (HMM) classification is crucial for patient care.
• Limited datasets and high annotation costs pose challenges in developing accurate HMM classification models.
• Active learning algorithms can optimize dataset selection to improve model training efficiency.

Purpose of the Study:

• To classify high myopic maculopathy (HMM) using limited datasets.
• To optimize the ALFA-Mix active learning algorithm for HMM classification.
• To minimize annotation costs while maintaining classification accuracy.

Main Methods:

• An optimized ALFA-Mix algorithm (ALFA-Mix+) was developed and compared against other algorithms.
• Experiments involved combining ALFA-Mix+ with deep learning models like ResNet18 and EfficientFormer.
• Performance was evaluated over 20 active learning rounds, selecting 100 images per round.

Main Results:

• ALFA-Mix+ demonstrated superior performance over other algorithms in terms of accuracy, sensitivity, specificity, and Kappa values.
• The EfficientFormer model achieved high performance on the HMM dataset (Accuracy: 0.8821).
• Combining ALFA-Mix+ with EfficientFormer yielded the best results (Accuracy: 0.8964, Sensitivity: 0.8643, Specificity: 0.9721, Kappa: 0.8537).

Conclusions:

• The ALFA-Mix+ algorithm effectively reduces the number of required samples without compromising classification accuracy.
• ALFA-Mix+ outperforms other algorithms in selecting valuable samples for training.
• The combination of ALFA-Mix+ and EfficientFormer significantly enhances HMM classification performance.