Researchers Track Age-related Macular Degeneration Progression in OCT Scans with Fusion CNN Networks

Age-related Macular Degeneration represents a significant threat to vision, and effective management relies on carefully tracking disease progression, particularly in response to anti-VEGF treatments. Philippe Zhang, Weili Jiang, and Yihao Li, along with Jing Zhang, Sarah Matta, and Yubo Tan, addressed this critical need by participating in the Monitoring Age-related Macular Degeneration Progression in Optical Coherence Tomography (MARIO) challenge. Their work focuses on automatically classifying changes observed in retinal scans and predicting future disease states, offering the potential to refine treatment strategies and improve patient outcomes. The team developed innovative approaches, including a fusion convolutional neural network and a patch progression masked autoencoder, achieving top-ten performance in both tasks of the challenge and demonstrating a promising advancement in automated disease monitoring.

Anti-vascular endothelial growth factor (anti-VEGF) treatments have demonstrated effectiveness in slowing the progression of neovascular AMD, and improved outcomes correlate with timely diagnosis and consistent monitoring. The team meticulously processed scans from 136 patients, collected using the Spectralis OCT device, dividing the data into training, validation, and testing sets. To ensure unbiased evaluation, organizers independently processed the validation and test datasets, maintaining the integrity of the challenge. A crucial initial step involved addressing inconsistencies in the OCT images, specifically vertical misalignment, using a dedicated image preprocessing pipeline called OCTIP.

Scientists trained several Feature Pyramid Network (FPN) architectures from the EfficientNet family to segment the relevant retinal regions, ultimately selecting FPN-EfficientNet-B6 and FPN-EfficientNet-B7 for optimal performance. By combining the outputs of these two models via a median filter, the team generated robust segmentation masks, effectively “flattening” the retina and aligning images along the depth axis. This preprocessing step eliminated irrelevant information, enabling a detailed analysis of the retinal layers. For the first task, classifying changes between consecutive OCT scans, the team explored both Early Fusion and Late Fusion Convolutional Neural Networks (CNNs).

The Early Fusion network concatenates paired OCT images, creating a fused input that then passes through a ResNet50 encoder, pre-trained on ImageNet, to extract robust feature maps. Scientists then employed fully connected layers to perform the final four-category classification: Reduced, Stable, Worsened, or Other. The team also investigated a Late Fusion network, which fuses features extracted from images at two different time points, providing an alternative approach to capturing temporal changes in the scans. Their work builds upon the established effectiveness of anti-VEGF treatments for neovascular AMD, which rely on early diagnosis and consistent monitoring to maximize patient outcomes. The first task involved classifying the evolution of AMD between pairs of consecutive OCT scans, categorizing progression as reduced, stable, worsened, or other. For the second task, researchers focused on predicting AMD progression over a three-month period using single OCT scans. Their approach involved a novel method for generating predicted future OCT scans and then classifying the evolution between the current scan and the generated prediction. This innovative technique, combined with the robust image preprocessing, allowed the team to accurately predict disease progression, achieving a top-10 ranking and demonstrating the potential for computer-assisted diagnostic tools to improve AMD management. The success of these methods highlights the power of deep learning and advanced image processing in advancing the field of ophthalmology and improving patient care.

Predicting AMD Progression with Deep Learning

The research team developed and evaluated new methods for tracking the progression of age-related macular degeneration (AMD) using optical coherence tomography (OCT) scans. Their approach involved two main tasks: classifying changes in OCT scans over time and predicting future scan appearances. For both tasks, the team employed convolutional neural networks, enhanced by a technique called OCTIP to improve image quality and feature extraction. Model ensembling further refined performance, combining the strengths of multiple models. The results demonstrate the potential of these advanced neural network architectures to improve the monitoring and treatment of AMD, achieving a Top 10 ranking in both tasks of the MARIO Challenge. However, the study acknowledges that predicting disease progression from a single scan remains difficult, and incorporating additional patient data, such as age and sex, did not improve accuracy. Future research should focus on refining image reconstruction techniques to generate more accurate predictions of future scan appearances, ultimately contributing to more personalized and effective treatment strategies for AMD.