An active learning method for diabetic retinopathy classification with uncertainty quantification.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Machine Learning

Background:

• Deep learning (DL) achieves state-of-the-art results in medical imaging but requires extensive annotated data and computational resources.
• The 'black-box' nature of DL models limits transparency and increases prediction uncertainty.
• Acquiring high-quality annotated medical data is challenging due to costs, limited expert annotators, and time constraints.

Purpose of the Study:

• To address the challenges of data scarcity, computational cost, and lack of transparency in deep learning for medical imaging.
• To propose a hybrid model integrating uncertainty quantification and active learning for efficient medical data annotation.
• To enhance the reliability and interpretability of deep learning models in medical diagnostic tasks.

Main Methods:

• A hybrid model combining a Bayesian convolutional neural network (BCNN) with an active learning strategy was developed.
• The BCNN was utilized as a feature descriptor to extract meaningful information from medical images.
• Active learning was employed to guide the annotation process, focusing on the most informative unlabeled data points.

Main Results:

• The proposed framework demonstrated state-of-the-art performance in diabetic retinopathy classification.
• The integration of BCNNs effectively quantified model uncertainty, providing insights into prediction reliability.
• The active learning approach significantly improved the efficiency of data annotation.

Conclusions:

• The hybrid BCNN and active learning model offers a robust solution for deep learning in medical imaging, overcoming key data and transparency challenges.
• This approach enhances the practical applicability of AI in healthcare by improving model interpretability and data efficiency.
• The framework shows significant potential for advancing automated medical image analysis and diagnosis.