A Comparative Study on Deep Convolutional Neural Networks and Histogram Equalization Techniques for Glaucoma Detection From Fundus Images

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. Read our paper 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

Model	Accuracy	Specificity	Sensitivity	F1 Score	Area Under ROC	Number of Parameters
VGG-16	0.9718	0.9516	0.9875	0.9753	0.9978	23104066
VGG-19	0.9789	0.9677	0.9875	0.9814	0.9933	28413762
ResNet-50	0.9577	0.9839	0.9375	0.9615	0.9956	57142914
ResNet-152	0.9507	0.9839	0.925	0.9548	0.9944	91926146
Inception v3	0.9085	0.8871	0.925	0.9193	0.9364	40677922
Xception	0.9296	0.9516	0.9125	0.9359	0.9794	54416682
DenseNet-121	0.9577	0.9516	0.9625	0.9625	0.9960	23815490
EfficientNetB7	0.9225	0.8871	0.95	0.9325	0.9625	106041497

Preprocessing Techniques Comparison

Two different preprocessing techniques were compared:

• Adaptive Histogram Equalisation
• Contrast-Limited Adaptive Histogram Equalisation

Results

I. Histogram Equalisation

Model	Accuracy	Specificity	Sensitivity	F1 Score	Area Under ROC	Number of Parameters
VGG-16	0.943661972	0.935483871	0.95	0.95	0.993548387	23104066
VGG-19	0.894366197	0.951612903	0.85	0.900662252	0.978427419	28413762
ResNet-50	0.936619718	0.951612903	0.925	0.942675159	0.991935484	57142914
ResNet-152	0.936619718	0.967741935	0.9125	0.941935484	0.98891129	91926146
Inception v3	0.767605634	0.870967742	0.6875	0.769230769	0.846774194	40677922
Xception	0.767605634	0.903225806	0.6625	0.762589928	0.898387097	54416682
DenseNet-121	0.929577465	0.887096774	0.9625	0.93902439	0.985080645	23815490
EfficientNetB7	0.838028169	0.822580645	0.85	0.855345912	0.926814516	106041497

II. Contrast-Limited Adaptive Histogram Equalisation

Model	Accuracy	Specificity	Sensitivity	F1 Score	Area Under ROC	Number of Parameters
VGG-16	0.922535211	1	0.8625	0.926174497	0.996572581	23104066
VGG-19	0.922535211	1	0.8625	0.926174497	0.996572581	28413762
ResNet-50	0.929577465	0.870967742	0.975	0.939759036	0.983770161	57142914
ResNet-152	0.957746479	0.935483871	0.975	0.962962963	0.994153226	91926146
Inception v3	0.809859155	0.774193548	0.8375	0.832298137	0.91733871	40677922
Xception	0.781690141	0.935483871	0.6625	0.773722628	0.921169355	54416682
DenseNet-121	0.950704225	0.935483871	0.9625	0.956521739	0.987701613	23815490
EfficientNetB7	0.894366197	0.838709677	0.9375	0.909090909	0.925604839	106041497

III. Comparing the Averages of Models