Deep-Learning-Based Visualization and Volumetric Analysis of Fluid Regions in Optical Coherence Tomography Scans.

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Accurate retinal fluid volume computation is crucial for diagnosing and monitoring retinal pathologies.
• Optical coherence tomography (OCT) is a key imaging modality for visualizing retinal structures.
• Existing methods for volumetric analysis can be labor-intensive and subjective.

Purpose of the Study:

• To develop and validate a deep learning-based visualization tool for quantifying fluid volume in OCT images.
• To assess the performance of different Class Activation Mapping (CAM) techniques for visualizing pathology-specific regions.
• To compare the accuracy of a standard Inception-ResNet-v2 model with a smaller variant for volumetric analysis.

Main Methods:

• Binary classification models (Inception-ResNet-v2) were developed for Intraretinal Fluid (IRF), Subretinal Fluid (SRF), and Pigmented Epithelial Detachment (PED).
• Multiple CAM techniques (Grad-CAM, Grad-CAM++, Score-CAM, Ablation-CAM, Self-Matching CAM) and a novel Ensemble-CAM were employed for visualization.
• A Graphical User Interface (GUI) integrated visualization heatmaps with region-growing and selective thresholding algorithms for volume computation.

Main Results:

• The deep learning tool accurately calculated fluid volumes in OCT scans, comparable to expert annotations.
• Ensemble-CAM provided robust visualization of pathology-specific regions.
• The smaller Inception-ResNet-v2 model demonstrated comparable performance with reduced computational resources.

Conclusions:

• Deep learning-based visualization techniques are relevant and useful for reliable volumetric analysis of retinal pathologies.
• The developed tool offers a promising approach for objective and efficient assessment of retinal fluid volume.
• This technology can aid in grading pathologies and evaluating treatment responses in clinical practice.