TLDR: Researchers developed deep learning models for the MARIO Challenge to classify Age-related Macular Degeneration (AMD) progression from OCT scans. For classifying evolution between two scans (Task 1), they used a Fusion CNN Network with ResNet50 and model ensembling, enhanced by OCTIP preprocessing. For predicting progression from a single scan (Task 2), they introduced the Patch Progression Masked Autoencoder (PPMAE) to predict a future OCT image, then used the Task 1 model for classification. Their methods achieved top-10 results, demonstrating improved accuracy and reconstruction performance, particularly with OCTIP.
Age-related Macular Degeneration (AMD) is a significant eye condition that impacts vision, particularly in individuals over 65. A specific type, neovascular AMD, can be managed with anti-vascular endothelial growth factor (anti-VEGF) treatments, but timely diagnosis and consistent monitoring are crucial for the best outcomes. Tracking how neovascular activity changes in Optical Coherence Tomography (OCT) scans is key to creating personalized and effective treatment plans for patients.
This critical area of research was the focus of the Monitoring Age-related Macular Degeneration Progression In Optical Coherence Tomography (MARIO) Challenge. A team of researchers, including Philippe Zhang and colleagues, participated in this challenge, presenting innovative deep learning solutions that achieved top-10 results in both tasks. You can find their full paper here: Patch Progression Masked Autoencoder with Fusion CNN Network for Classifying Evolution Between Two Pairs of 2D OCT Slices.
Understanding the MARIO Challenge
The MARIO Challenge was designed to evaluate algorithms for detecting changes in neovascular activity from OCT scans. It comprised two main tasks:
- Task 1: Classifying the evolution between two pairs of 2D slices from consecutive OCT acquisitions. This involved categorizing the condition as Reduced, Stable, Worsened, or Other.
- Task 2: Predicting AMD progression over the next three months based on current exam data, classifying it into Reduced, Stable, or Worsened categories.
Innovative Solutions for Task 1: Fusion CNN Network
For Task 1, the team developed a Fusion CNN Network, leveraging ResNet50 as the backbone encoder. They explored two main fusion approaches:
- Early Fusion: This method combines the two OCT images (from different time points) at the input level, concatenating them along the channel dimension before feeding them into the ResNet50 encoder.
- Late Fusion: In this approach, features are extracted independently from each OCT image using separate ResNet50 encoders. These extracted features are then concatenated and passed through fully connected layers for classification.
To further enhance performance and robustness, the team adopted a 4-fold cross-validation strategy, training multiple models and combining their results through ensembling. A crucial preprocessing step, Optical Coherence Tomography Image Preprocessing (OCTIP), was also applied. OCTIP helps reduce noise, align images vertically, and zoom in on the retina, providing cleaner and more consistent data for the neural network.
Advanced Prediction for Task 2: Patch Progression Masked Autoencoder (PPMAE)
Task 2 presented a unique challenge: predicting disease progression from a single OCT scan. To overcome this, the researchers introduced the Patch Progression Masked Autoencoder (PPMAE). Inspired by traditional Masked Autoencoders, PPMAE takes a current OCT image (at time t0), masks a significant portion (75%) of it, and then learns to predict the corresponding patches of a future OCT image (at time t1).
This innovative approach allows the model to capture temporal changes and learn how the disease progresses over time. Once the future OCT image (t1) is predicted, the classification model developed for Task 1 is then used to classify the evolution between the current (t0) and the generated future (t1) scans, effectively turning a single-scan prediction problem into a two-scan classification problem.
Key Role of OCTIP and Promising Results
The OCTIP preprocessing technique proved to be vital for both tasks. It significantly improved image quality, enhanced classification accuracy, and enabled more precise feature extraction, which also contributed to better image reconstruction in Task 2. The results were impressive: OCTIP improved F1 scores and rank correlation for both Early and Late Fusion methods in Task 1. For Task 2, the PPMAE model achieved significantly lower Mean Squared Error (MSE) in reconstructing future images compared to traditional methods, especially when combined with OCTIP. This better reconstruction directly led to improved classification performance, with the PPMAE-based method achieving strong F1 scores and other metrics.
Also Read:
- Precision Retinal Image Registration: The GPO Framework for Enhanced Alignment
- Advanced AI Model Enhances Brain Tumor Detection in MRI Scans
Conclusion and Future Directions
The solutions presented by Philippe Zhang and his team for the MARIO Challenge demonstrate the significant potential of advanced neural network architectures in developing more personalized and effective treatment strategies for AMD. While predicting disease progression from a single exam remains complex, the PPMAE offers a promising avenue. Future research will focus on further improving image reconstruction methods to generate even better future images, which is essential for continued progress in this field.
