Two-Stage Decoupling Framework for Variable-Length Glaucoma Prognosis

Area of Science:

• Ophthalmology
• Medical Informatics
• Artificial Intelligence

Background:

• Glaucoma is a primary cause of irreversible blindness globally, necessitating accurate prognosis for timely intervention.
• Current glaucoma prognosis methods face limitations due to fixed-length input constraints and challenges with small, heterogeneous datasets.
• Existing end-to-end models struggle with data scarcity common in glaucoma research.

Purpose of the Study:

• To develop a flexible and robust framework for glaucoma prognosis that accommodates variable-length sequential data.
• To overcome limitations of fixed-length inputs and small dataset sizes in traditional glaucoma prediction models.
• To improve the accuracy and efficiency of identifying patients at risk of glaucoma progression.

Main Methods:

• Proposed a Two-Stage Decoupling Framework (TSDF) for variable-length glaucoma prognosis.
• Stage 1: Utilized self-supervised learning for feature representation, aggregating diverse glaucoma datasets regardless of supervisory information.
• Stage 2: Implemented an attention-based temporal aggregation module to process variable-length sequential inputs efficiently.

Main Results:

• The TSDF demonstrated significant enhancement in model performance across two benchmark datasets (OHTS and GRAPE).
• The framework proved effective and robust, even with datasets differing significantly in scale and clinical settings.
• The approach maintained a compact parameter size while maximizing the utilization of all available patient data.

Conclusions:

• The Two-Stage Decoupling Framework (TSDF) offers a flexible and effective solution for glaucoma prognosis using variable-length data.
• The proposed method successfully addresses challenges posed by data heterogeneity and size limitations in glaucoma research.
• This approach holds promise for improving early detection and intervention strategies to prevent glaucoma-induced blindness.