MARIO Challenge – Team DF41

This repository contains our team solution for the MICCAI 2024 MARIO Challenge, focusing on retinal disease progression analysis from OCT imaging.

This work is based on our paper:

Patch Progression Masked Autoencoder with Fusion CNN Network for Classifying Evolution Between Two Pairs of 2D OCT Slices
P. Zhang et al., MICCAI 2024

🔗 Read the paper

---

🧠 Overview

Age-related Macular Degeneration (AMD) is a major cause of vision loss. Monitoring disease progression using OCT imaging is critical for treatment decisions.

The MARIO challenge addresses two tasks:

• Task 1: Classify disease evolution between two consecutive OCT scans
• Task 2: Predict disease progression from a single OCT scan

Our approach combines:

• CNN-based fusion models for classification
• Masked Autoencoders (MAE) for temporal prediction
• Model ensembling for robustness
• OCT preprocessing (OCTIP) for alignment and ROI extraction

---

📂 Repository Structure

.
├── configs/                  # YAML configuration files
├── csv/                      # Input metadata (challenge format)
├── models/                   # Model definitions (weights not included)
├── utils/                    # Preprocessing, datasets, MAE, scoring
├── inference_pipeline_task_1.py
├── inference_pipeline_task_2.py
├── Dockerfile
├── requirements.txt
└── README.md

---

⚙️ Methodology

🔹 Task 1 – Evolution Classification

We model disease evolution using fusion CNN architectures.

• Early Fusion: concatenate (t0, t1) along channels → ResNet50 → 4-class output
• Late Fusion: extract features independently (2048 + 2048) → FC classifier
• Training: 4-fold cross-validation
• Inference: ensemble across folds

---

🔹 Task 2 – Progression Prediction

Only one OCT scan (t0) is available. We convert it into a temporal problem.

🧩 Patch Progression Masked Autoencoder (PPMAE)

We employed a Masked Autoencoder to generate a synthetic OCT image. Instead of reconstructing the same image, we predict the future OCT (t1) from t0.

• Mask ~75% of patches
• Encode visible patches
• Decode to reconstruct future patches
• Loss: MSE on predicted t1 patches

---

🔁 Final Pipeline (Task 2)

1. Input OCT at time t0
2. Generate predicted OCT t1 using PPMAE
3. Apply Task 1 classifier on (t0, predicted t1)

👉 Converts Task 2 into a synthetic temporal classification task

---

🧪 Data Processing (OCTIP)

• Retina alignment (flattening)
• Noise removal
• Segmentation (EfficientNet-based FPN)
• ROI-focused crops

---

📊 Results

• Top 10 – MICCAI 2024 MARIO Challenge
• Gains from:
  • OCTIP preprocessing
  • Ensembling
  • PPMAE-based reconstruction

Qualitative examples (PPMAE):

---

🚀 Usage

Task 1

python inference_pipeline_task_1.py

Task 2

python inference_pipeline_task_2.py

---

⚠️ Data & Models

This repository does not include:

• Challenge dataset
• Model checkpoints (.pth / .hdf5)

To run the code, provide:

• MARIO dataset
• Pretrained weights (paths configurable via YAML)

---

🧰 Configuration

• configs/config_task1.yaml
• configs/config_task2.yaml

---

🧑‍💻 Author

Philippe Zhang
PhD – Medical Imaging & Deep Learning

---

🔗 References

• MARIO Challenge: https://youvenz.github.io/MARIO_challenge.github.io/
• OCTIP: https://github.com/leto-atreides-2/octip