OCTMNIST Retinal Classification

Deep learning project focused on multiclass retinal OCT image classification using a ResNet-based pipeline.

Overview

This repository documents a computer vision project built on the MedMNIST OCTMNIST dataset. The goal was to compare training settings and augmentation strategies for retinal OCT classification, then identify a strong-performing configuration based on empirical results.

Rather than treating the task as a single training run, this project was organized as a series of controlled experiments across:

• image augmentation strategies,
• batch sizes,
• training epochs, and
• learning rates.

The repository includes experiment scripts, result summaries, and supporting figures used to compare model configurations.

Task and Dataset

• Task: Multiclass retinal OCT image classification
• Dataset: OCTMNIST from MedMNIST
• Framework: PyTorch
• Selected backbone: ResNet-50

OCT imaging is widely used in retinal assessment, making this a useful benchmark for image classification, experimental model tuning, and medical imaging workflows.

Experimental Focus

The project evaluates how model performance changes under different training conditions and backbone choices.

Augmentation Comparison

The experiments compared several image transformation strategies, including:

• baseline/original input,
• AugMix,
• colour jitter,
• increased sharpness, and
• automatic contrast adjustment.

Across these runs, AugMix gave the most balanced overall performance and was selected for the final configuration.

Hyperparameter Exploration

The experiments also evaluated the effect of:

• batch size,
• number of epochs, and
• learning rate

on classification performance. The figure below summarizes the main trends used to choose the final configuration.

Model Comparison

The project also compared ResNet-18 and ResNet-50 as candidate backbones. ResNet-50 showed slightly stronger AUC and precision, while ResNet-18 had slightly higher recall. Based on that tradeoff, ResNet-50 was selected as the final backbone.

Final Selected Configuration

The final selected configuration used:

• Architecture: ResNet-50
• Augmentation: AugMix
• Batch size: 32
• Epochs: 50
• Learning rate: 1e-4

This configuration provided a strong tradeoff between predictive performance and training setup.

Reported Results

Final test-set results from the selected experiment:

• AUC: 95.14%
• Accuracy: 74.50%
• Precision: 98.99%
• Recall: 80.24%

Repository Structure

.
├── code/                 # Training scripts and experiment variants
├── images/               # Figures used in the README
├── Evaluation metrics.xlsx
└── README.md

Acknowledgement

This project uses the MedMNIST / OCTMNIST benchmark dataset for experimentation and evaluation.