Retinal OCT Disease Classification using Deep Learning

This project performs multi-class classification of retinal diseases using Optical Coherence Tomography (OCT) images. The model is trained to classify images into four categories: CNV, DME, DRUSEN, and NORMAL using transfer learning techniques.

Dataset

The dataset used is the OCT2017 retinal dataset, which contains 4 main classes:

• CNV (Choroidal Neovascularization)
• DME (Diabetic Macular Edema)
• DRUSEN
• NORMAL

The data is organized into train/, test/, and val/ folders, each containing class-based subfolders.

Model Architecture

We used MobileNetV2 (pre-trained on ImageNet) as a feature extractor. The top layers were replaced with a custom classification head:

• GlobalAveragePooling2D
• Dense(128, relu)
• Dropout(0.2)
• Dense(4, softmax)

The feature extractor is frozen during initial training.

Techniques Used

• ImageDataGenerator for real-time data augmentation
• class_weight handling for imbalanced classes
• Transfer Learning with MobileNetV2
• Callbacks: EarlyStopping, ReduceLROnPlateau, ModelCheckpoint

Evaluation Metrics

• Accuracy
• Precision
• Recall
• F1-Score
• Confusion Matrix

Best performance:

Accuracy:       93%
Precision avg:  0.93
Recall avg:     0.93
F1-score avg:   0.93

Results

The model performed well across all four classes, with minor confusion between CNV and DRUSEN.

Confusion Matrix:

	CNV	DME	DRUSEN	NORMAL
CNV	217	7	18	0
DME	2	223	4	13
DRUSEN	10	7	221	4
NORMAL	0	3	4	235

Sample Misclassification:

Displayed misclassified images with predicted and true labels for manual analysis.

Libraries Used

• TensorFlow / Keras
• OpenCV
• NumPy, Pandas
• Scikit-learn
• Seaborn / Matplotlib