A Semi-supervised Reinforcement Learning Framework Incorporating the Multi-scale IncepMambaNet Network for Glaucoma...

Area of Science:

• Ophthalmology
• Artificial Intelligence
• Machine Learning

Background:

• Glaucoma is a primary cause of irreversible blindness globally.
• Current AI models for glaucoma progression prediction using OCT and visual field testing face challenges due to limited labeled data and ambiguous structural differences.
• These limitations hinder the development of robust AI diagnostic tools.

Purpose of the Study:

• To propose a novel glaucoma progression prediction method using a semi-supervised reinforcement learning (SSRL) framework.
• To address the bottlenecks of limited labeled data and ambiguous structural differences in glaucoma.
• To enhance the generalization ability and accuracy of AI models in glaucoma prediction.

Main Methods:

• Developed a semi-supervised reinforcement learning (SSRL) framework incorporating a key experience filtering strategy (KEFS) to prioritize high-value training samples and optimize pseudo-labels.
• Introduced an entropy regularization technique to prevent premature model convergence and improve unlabeled data exploration.
• Proposed the IncepMambaNet multi-scale network, integrating the VMamba model's visual state space (VSS) module with CNNs for improved feature extraction.

Main Results:

• The SSRL model demonstrated significant performance improvements over traditional supervised learning methods for glaucoma prediction.
• IncepMambaNet achieved leading performance compared to state-of-the-art (SOTA) networks, with notable gains in macro-F1 (0.8%), F2-score (8.8%), and AUC (0.9%).
• Ablation studies confirmed the effectiveness of the multi-branch structure and channel reordering in the IncepMambaNet architecture.

Conclusions:

• The proposed SSRL framework with IncepMambaNet offers a powerful approach to glaucoma progression prediction, effectively addressing data limitations.
• The IncepMambaNet architecture exhibits superior feature capture and generalization capabilities, outperforming existing SOTA models.
• This research advances AI applications in ophthalmology, paving the way for more accurate and reliable glaucoma diagnosis.