Simultaneous Segmentation of Geographic Atrophy in Longitudinally Acquired Fundus Autofluorescence Images

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Measuring geographic atrophy (GA) progression in clinical trials is challenging due to its slow and variable nature.
• Accurate GA measurement is crucial for evaluating treatment efficacy in clinical trials.

Purpose of the Study:

• To develop and validate deep learning algorithms for accurate and robust measurement of GA growth rates.
• To improve the automated annotation of pre-existing and newly expanding GA regions using fundus autofluorescence (FAF) image pairs.

Main Methods:

• A DeepLabV3+ (DL) segmentation model was enhanced to simultaneously process longitudinally acquired, spatially aligned FAF image pairs (DL_JOINT).
• The model (DL_JOINT+SCL) was further modified to ensure longitudinal consistency of existing GA pixels.
• Performance was evaluated against manual expert annotations using metrics like precision, recall, Dice coefficient, and correlation for GA area and growth rates.

Main Results:

• The DL_JOINT+SCL model achieved high performance: precision=0.86, recall=0.92, Dice=0.89, with strong correlations for GA area (R²=0.97) and growth rate (R²=0.82).
• Simultaneous segmentation of image pairs significantly improved growth rate correlation compared to a static model (R²=0.75).
• Image registration for spatial alignment was critical, enhancing accuracy by mapping regions to a common coordinate system.

Conclusions:

• Simultaneous segmentation of longitudinal FAF images improves the accuracy and robustness of GA progression measurements.
• Maintaining longitudinal consistency in measurements is vital for the translation of automated algorithms in clinical trials.
• These advancements minimize measurement variability, enabling more effective assessment of clinical trial outcomes in GA.