Deep CNNs and Histogram Equalization for Glaucoma Detection

Using CNNs to classify an image into normal or glaucomatous, using retinal fundus images by transfer learning, and evaluating 2 histogram equalisation-based image preprocessing techniques. The dataset used is ACRIMA, containing 705 labelled images: 396 glaucomatous images and 309 normal images.

This project was a collaboration with H Shafeeq Ahmed. Our preprint is hosted on medRxiv here, you can also read it here.

Model Comparison

The CNN models were fit using 70% of the dataset for training, 10% for validation and 20% for testing.

The model architecture used is as follows:

• An input layer (256, 256, 3)
• A data augmentation layer
• The base model with image-net weights
• A flatten layer
• A dense layer with ReLU activation
• A 0.5 dropout layer
• A dense output layer with softmax activation

Data augmentation involved random change in contrast, flip along horizontal or vertical, rotation, and translation of the images:

The notebooks can be accessed in the notebooks folder folder of this repo. Here is an example of the notebook used for training a model based on VGG-19 using CLAHE for image preprocessing.

Results

Preprocessing Techniques Comparison

Two different preprocessing techniques were compared:

• Adaptive Histogram Equalisation
• Contrast-Limited Adaptive Histogram Equalisation

Results

I. Histogram Equalisation

II. Contrast-Limited Adaptive Histogram Equalisation

III. Comparing the Averages of Models