An integrated time adaptive geographic atrophy prediction model for SD-OCT images

Area of Science:

• Ophthalmology
• Artificial Intelligence
• Medical Image Analysis

Background:

• Geographic atrophy (GA) lesion growth prediction is crucial for patient management and treatment assessment.
• Current methods may lack accuracy in predicting the dynamic progression of GA.
• Automated prediction models can aid ophthalmologists in understanding disease trajectory.

Purpose of the Study:

• To develop an integrated, time-adaptive prediction model for identifying future geographic atrophy (GA) growth locations.
• To enhance the accuracy and reliability of GA progression prediction using advanced AI techniques.
• To evaluate the model's performance across various scenarios, including data similarity, temporal factors, and generalization.

Main Methods:

• Developed a hybrid model combining a bi-directional long short-term memory (BiLSTM) network for prediction and a convolutional neural network (CNN) for refinement.
• Integrated time factors into the BiLSTM module to account for irregular time intervals between follow-up visits.
• Employed 10 distinct scenarios to rigorously evaluate prediction accuracy, data importance, temporal effects, and model generalization.

Main Results:

• The model achieved high average Dice Indexes (DI) across scenarios (0.86-0.92), indicating accurate prediction of GA regions.
• Integrating time factors improved prediction accuracy by approximately 10%.
• The CNN refinement effectively removed false positive predictions and enhanced overall accuracy, with 2 sequential visits outperforming single visits for training.

Conclusions:

• The proposed time-adaptive prediction model demonstrates feasibility and effectiveness for predicting geographic atrophy (GA) growth.
• The model shows good generalization performance across different patient regions.
• Prior information similarity and temporal dynamics significantly influence prediction accuracy, highlighting the model's robustness.