A fundus image classification framework for learning with noisy labels.

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Fundus image analysis is crucial for diagnosing eye diseases.
• Supervised deep learning models require high-quality, accurately labeled data.
• Noisy labels in fundus datasets significantly impair the performance of these models.

Purpose of the Study:

• To develop a robust noisy label learning framework for multiclass classification of fundus diseases.
• To enhance the performance of computer-aided diagnosis systems for eye conditions despite label noise.
• To improve the generalization capabilities of deep learning models in medical image analysis.

Main Methods:

• A framework combining data cleansing (DC), adaptive negative learning (ANL), and sharpness-aware minimization (SAM) was proposed.
• The DC module filters noisy labels using prediction confidence.
• ANL modifies the loss function with complementary labels, and SAM optimizes loss and its sharpness for better generalization.

Main Results:

• The proposed framework demonstrated significant improvements in classifying fundus diseases with noisy labels.
• Experiments on private and public datasets confirmed the method's effectiveness.
• The combined approach successfully mitigated the negative impact of label noise on diagnostic performance.

Conclusions:

• The developed noisy label learning framework effectively enhances the accuracy of fundus disease classification.
• The integration of DC, ANL, and SAM modules offers a robust solution for handling imperfectly labeled medical image data.
• This approach holds promise for improving computer-aided diagnosis systems in ophthalmology.