🩺 Retinal OCT Segmentation with Shallow CNN, ResNet & Graph Cut

📌 Overview

This project performs retinal OCT (Optical Coherence Tomography) image segmentation using the Duke OCT dataset. We focus on two conditions: Normal and Age-related Macular Degeneration (AMD).

Dataset available @ https://people.duke.edu/~sf59/RPEDC_Ophth_2013_dataset.htm

Methods implemented:

• Mask Generation – create ground-truth masks from .mat files
• Data Preparation – splitting data into train/val/test and preprocessing
• Shallow CNN – custom lightweight encoder–decoder network
• ResNet34 Segmentation – deep encoder–decoder with transfer learning
• Graph Cut Refinement – post-processing to smooth CNN predictions

---

⚙️ Workflow

1. Mask Generation (Mask_generation.ipynb)

   • Converts Duke .mat files into multiclass masks (4 classes, values 0–3).
   • Saves both visualization PNGs and raw .npy masks.

2. Data Preparation (data_preparation.ipynb)

   • Resizes all images/masks to 256×256
   • Splits data: 70% Train, 15% Validation, 15% Test

3. Model Training

   • Shallow CNN: custom 3-block encoder–decoder

   • ResNet34: pretrained encoder with custom decoder

   • Training settings (both):

     • BATCH_SIZE = 8
     • EPOCHS = 20
     • LR = 1e-4
     • Loss = CrossEntropy, Optimizer = Adam

4. Prediction & Graph Cut

   • Generate masks from trained models

   • Apply Graph Cut (α-expansion) with:

     • smooth_weight = 10
     • max_iter = 5

   • Produces smoother, more accurate boundaries

5. Evaluation Metrics

   • Dice Score (per-class + mean)
   • IoU (per-class + mean)
   • Precision & Recall (per-class + mean)

---

📊 Results

Training and evaluation are included inside the ResNet (AMD/Normal) and Shallow CNN (AMD/Normal) notebooks. They report Dice, IoU, Precision, and Recall for each experiment on the test set.

👉 After training, you can run the included evaluation functions to generate a table like this:

Model	Dataset	Dice (mean)	IoU (mean)	Precision	Recall
Shallow CNN + GC	AMD	0.8067	0.7156	0.8423	0.8007
Shallow CNN + GC	Normal	0.7505	0.6385	0.7763	0.7586
ResNet34 + GC	AMD	0.9238	0.8792	0.9257	0.9235
ResNet34 + GC	Normal	0.9149	0.8683	0.9101	0.9232

ResNet34 + Graph Cut on AMD Dataset

Shallow CNN + Graph Cut on AMD Dataset