A novel approach to macular edema detection: DeepLabv3+ segmentation and VGG with vision transformer classification

Figures & data

Figure 1. Sample fundus images.

Figure 2. Overview of suggested work.

Table

Pseudocode for Data pre-processing

Input: image- fundus image

Output: Augmented image

function CLAHE(image, clip limit) // Contrast Enhancement

grayscale_image = convert_to_grayscale(image)

clahe = create_CLAHE(clip Limit = cliplimit)

enhanced_image = clahe.apply(grayscale_image)

return enhanced_image

function illumination_correction(image, desired_intensity)

// Illumination Correction

local_average_intensity = calculate_local_average_intensity

(image)

correction_factor = desired_intensity / local_average_intensity

corrected_image = image * correction_factor

return corrected_image

set: augmented_images = []

For image in images do // Cropping

cropped_image = crop_image(image)

augmented_images.append(cropped_image)

For angle in [90, 120, 180, 270] do // Rotation

rotated_image = rotate_image(image, angle)

augmented_images.append(rotated_image)

End For

horizontally_flipped_image = flip_image_horizontally

(image)

vertically_flipped_image = flip_image_vertically(image)

augmented_images.extend([horizontally_flipped_image,

vertically_flipped_image])

return augmented_images

Figure 3. Segmentation using Deeplabv3+.

Figure 4. Architecture of VGG-16 and VGG-19 for feature extraction.

Table 1. Results of classification evaluation using the confusion matrix parameters and three distinct learning rates.

	Learning rate	0.00001	0.0001	0.001
VGG-19	Accuracy	99.67	99.53	99.00
	sensitivity	99.65	99.25	98.38
	specificity	99.87	99.67	99.54
	positive predictive value	99.35	99.27	98.67
	F1 score	99.52	99.32	98.07
VGG-16	Accuracy	98.53	98.47	98.17
	sensitivity	98.29	99.15	98.38
	specificity	98.47	99.53	99.37
	positive predictive value	98.25	99.23	98.18
	F1 score	98.47	99.29	98.00

Table 2. Performance comparison between non-segmented and segmented images.

	Parameter	Segmented imaged	Non-segmentation images
VGG-19	Accuracy (%)	99.53	99.47
	Sensitivity (%)	99.37	99.28
	Specificity (%)	99.68	99.62
	positive predictive value (%)	99.27	99.12
	F1 score (%)	99.26	99.18
	Training time (second)	2.68 × 105	2.75 × 105
	Testing time (second)	28.24	33.14
VGG-16	Accuracy (%)	98.34	98.23
	Sensitivity (%)	98.78	98.65
	Specificity (%)	98.63	99.58
	positive predictive value (%)	98.45	98.32
	F1 score (%)	98.40	98.27
	Training time (second)	2.72 × 105	2.85 × 105
	Testing time (second)	30.47	37.48

Table 3. Comparison of number of Parameters of VGG-16 and VGG-19.

Model	Parameters (M)
VGG-16	14.71
VGG-19	20.02

Figure 5. MSE curve of (a) VGG-19 and (b) VGG-16.

Figure 6. MAE curve of (a) VGG-19 and (b) VGG-16.

Figure 7. The loss during the training and validation stages for VGG-16 and VGG-19.

Table 4. Performance of the segmentation algorithm.

Segmentation algorithms	DSC	J	Kappa	MeanIoU
DeepLabv3+	98.54	97.46	96.36	95.93
ResNet-50	94.97	95.37	94.93	94.28
MobileNetV2	96.66	93.78	95.48	95.37
Inception v3	93.57	92.39	93.34	94.48
ResNext50	95.54	94.57	94.76	94.75

Figure 8. Confusion matrix. a) VGG-19; b) VGG-16.

Figure 9. (a) ROC Curve of VGG-19 and (b) ROC curve of VGG-16.

Data availability statement

Data sharing is applicable to this article as a publicly available dataset analyzed during the current study