Rapid Retinopathy Detection using Ablation-Guided Deep Learning

M S  Rahman; Khondoker Shaila Sharmin

doi:10.12856/JHIA-2026-v13-i1-615

Rapid Retinopathy Detection using Ablation-Guided Deep Learning

M S Rahman
, Khondoker Shaila Sharmin (Contributor)

School of Computing, Eng & Intel. Sys

Research output: Contribution to journal › Article › peer-review

3 Downloads (Pure)

Abstract

Retinal fundus images have the challenge of differentiating the stages of diabetic retinopathy, often causing confusion between proliferative retinopathy to severe retinopathy or no retinopathy to mild retinopathy. Without developing complex technologies, human accuracy differed across image datasets. By using a deep learning CNN as the feature extractor and machine learning classifiers, this study achieved a remarkable accuracy of 91%. The ablation study utilised VGG16, Resnet152V2, and Xception as feature extractors together with Random Forests and K-Nearest Neighbour (KNN) classifiers. The best performance was achieved with Xception as the feature extractor and as the classifier Random Forest.

Original language	English
Article number	615
Pages (from-to)	17-31
Number of pages	15
Journal	The Journal of Health Informatics in Africa
Volume	13
Issue number	1
Early online date	28 Feb 2026
DOIs	https://doi.org/10.12856/JHIA-2026-v13-i1-615
Publication status	Published (in print/issue) - 28 Feb 2026

Bibliographical note

UN SDGs

This output contributes to the following UN Sustainable Development Goals (SDGs)

SDG 3 Good Health and Well-being

Keywords

Diabetic Retinopathy/diagnosis
Deep Learning
Ablation study
convolutional neural networks (CNN)
transfer learning
random forest (RF)

Access to Document

10.12856/JHIA-2026-v13-i1-615Licence: CC BY-NC

615Final published version, 1.89 MBLicence: CC BY-NC

Cite this

@article{9fc1b252cd544c9686a05fcc6ed87188,

title = "Rapid Retinopathy Detection using Ablation-Guided Deep Learning",

abstract = "Retinal fundus images have the challenge of differentiating the stages of diabetic retinopathy, often causing confusion between proliferative retinopathy to severe retinopathy or no retinopathy to mild retinopathy. Without developing complex technologies, human accuracy differed across image datasets. By using a deep learning CNN as the feature extractor and machine learning classifiers, this study achieved a remarkable accuracy of 91\%. The ablation study utilised VGG16, Resnet152V2, and Xception as feature extractors together with Random Forests and K-Nearest Neighbour (KNN) classifiers. The best performance was achieved with Xception as the feature extractor and as the classifier Random Forest.",

keywords = "Diabetic Retinopathy/diagnosis, Deep Learning, Ablation study, convolutional neural networks (CNN), transfer learning, random forest (RF)",

author = "Rahman, \{M S\} and Sharmin, \{Khondoker Shaila\}",

note = "{\textcopyright} 2026 JHIA.",

year = "2026",

month = feb,

day = "28",

doi = "10.12856/JHIA-2026-v13-i1-615",

language = "English",

volume = "13",

pages = "17--31",

journal = "The Journal of Health Informatics in Africa",

issn = "2197-6902",

number = "1",

}

TY - JOUR

T1 - Rapid Retinopathy Detection using Ablation-Guided Deep Learning

AU - Rahman, M S

A2 - Sharmin, Khondoker Shaila

PY - 2026/2/28

Y1 - 2026/2/28

N2 - Retinal fundus images have the challenge of differentiating the stages of diabetic retinopathy, often causing confusion between proliferative retinopathy to severe retinopathy or no retinopathy to mild retinopathy. Without developing complex technologies, human accuracy differed across image datasets. By using a deep learning CNN as the feature extractor and machine learning classifiers, this study achieved a remarkable accuracy of 91%. The ablation study utilised VGG16, Resnet152V2, and Xception as feature extractors together with Random Forests and K-Nearest Neighbour (KNN) classifiers. The best performance was achieved with Xception as the feature extractor and as the classifier Random Forest.

AB - Retinal fundus images have the challenge of differentiating the stages of diabetic retinopathy, often causing confusion between proliferative retinopathy to severe retinopathy or no retinopathy to mild retinopathy. Without developing complex technologies, human accuracy differed across image datasets. By using a deep learning CNN as the feature extractor and machine learning classifiers, this study achieved a remarkable accuracy of 91%. The ablation study utilised VGG16, Resnet152V2, and Xception as feature extractors together with Random Forests and K-Nearest Neighbour (KNN) classifiers. The best performance was achieved with Xception as the feature extractor and as the classifier Random Forest.

KW - Diabetic Retinopathy/diagnosis

KW - Deep Learning

KW - Ablation study

KW - convolutional neural networks (CNN)

KW - transfer learning

KW - random forest (RF)

U2 - 10.12856/JHIA-2026-v13-i1-615

DO - 10.12856/JHIA-2026-v13-i1-615

M3 - Article

SN - 2197-6902

VL - 13

SP - 17

EP - 31

JO - The Journal of Health Informatics in Africa

JF - The Journal of Health Informatics in Africa

IS - 1

M1 - 615

ER -

Rapid Retinopathy Detection using Ablation-Guided Deep Learning

Abstract

Bibliographical note

UN SDGs

Keywords

Access to Document

Fingerprint

Cite this