Foundation Model With Uncertainty Estimation-Based Active Learning for Retinal Image Classification.

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Automated retinal disease diagnosis is hindered by the extensive expert annotation required for training machine learning models.
• Current active learning methods face limitations in efficiently selecting informative samples for training, especially in low-data regimes.

Purpose of the Study:

• To develop and evaluate an active learning framework using a Foundation Model with Uncertainty Estimation (FMUE) for efficient retinal image annotation.
• To integrate evidential uncertainty estimation for improved sample selection in retinal diagnosis across Optical Coherence Tomography (OCT) and Color Fundus Photography (CFP) modalities.

Main Methods:

• Developed an FMUE-based active learning framework incorporating evidential deep learning for uncertainty estimation.
• Guided sample selection using an uncertainty-aware classifier to prioritize informative data points.
• Evaluated the framework's performance on four retinal imaging datasets, comparing it against traditional active learning methods like entropy-based selection and Bayesian Active Learning by Disagreement (BALD).

Main Results:

• The FMUE framework demonstrated superior performance over traditional methods, achieving accuracy improvements of 0.249 for CFP and 0.194 for OCT with only 2-4% annotation data.
• Achieved a sample selection speed 9 times faster than BALD in specific environments.
• Evidential uncertainty guidance led to more balanced category distribution and better identification of underrepresented retinal diseases.

Conclusions:

• Combining foundation models with evidential uncertainty estimation effectively addresses annotation challenges in retinal imaging.
• The proposed framework offers practical clinical advantages through enhanced sample selection and computational efficiency.
• This approach facilitates the deployment of automated diagnostic systems for retinal diseases.