Multi-Scale Spatio-Temporal Transformer-Based Imbalanced Longitudinal Learning for Glaucoma Forecasting From...

Area of Science:

• Ophthalmology
• Artificial Intelligence
• Medical Imaging

Background:

• Glaucoma is a leading cause of irreversible blindness, necessitating early detection and intervention.
• Forecasting glaucoma using sequential fundus images offers a promising approach for proactive patient management.
• Existing methods face challenges with irregularly sampled data and imbalanced class distributions.

Purpose of the Study:

• To develop an advanced deep learning model for accurate glaucoma forecasting from sequential fundus images.
• To address the limitations of irregular data sampling and class imbalance in disease prediction models.
• To evaluate the model's performance and generalization capabilities on medical datasets.

Main Methods:

• Introduction of the Multi-scale Spatio-temporal Transformer Network (MST-former) utilizing a transformer architecture for sequential image analysis.
• Implementation of a multi-scale structure for comprehensive spatial feature extraction at various resolutions.
• Design of a time distance matrix to handle non-linearly scaled temporal attention for irregularly sampled data.
• Application of a temperature-controlled Balanced Softmax Cross-entropy loss function to mitigate class imbalance issues.

Main Results:

• The MST-former achieved an Area Under the Curve (AUC) of 96.6% for glaucoma forecasting on the SIGF dataset.
• Demonstrated strong generalization by achieving 88.2% accuracy in predicting mild cognitive impairment and Alzheimer's disease on the ADNI MRI dataset.
• Ablation studies confirmed the effectiveness of individual components in handling data irregularities and class imbalance.

Conclusions:

• The proposed MST-former method significantly advances glaucoma forecasting accuracy and reliability.
• The model's ability to handle complex data challenges highlights its potential for broad application in medical image analysis.
• This approach offers a robust tool for early screening and intervention of eye diseases and potentially other neurological conditions.