Introduction
A comparative deep-learning study of four convolutional neural network architectures for classifying diabetic retinopathy severity from retinal fundus images.
Back to Projects
Deep Learning · Medical Imaging · Case Study
Automated retinal fundus image severity classification
Role
Deep Learning Team
Core Stack
Python · TensorFlow · PyTorch
Model Comparison
4 CNN Architectures
Status
In Progress
The Classification Problem
Five grades. One retinal image.
Manual grading of retinal images is time-intensive and depends on access to eye-care specialists. This project investigates automated classification across five diabetic retinopathy severity stages. The severity labels follow an ordered scale, so evaluation considers how far a prediction is from the reference grade.
Image-to-grade workflow
Research pipelineAPTOS Dataset
01 · Image data
Start with retinal fundus images.
The study uses the APTOS 2019 Blindness Detection training set: approximately 3,660 labeled RGB retinal images across five severity grades.
Source
APTOS 2019
Image type
RGB fundus
Dataset access
Download separately
Dataset files are not included in the repository; access is subject to Kaggle competition terms.
1 of 4 research steps
Architecture Comparison
Four approaches, shared conditions.
TensorFlow · PyTorch
01 · Baseline · Trained from scratch
Custom CNN
A from-scratch convolutional model establishes a baseline for comparing the value of transfer learning.
Baseline role
Shows how a model trained from scratch compares with ImageNet-initialized architectures.
02 · Transfer learning · ImageNet
DenseNet121
Dense connectivity reuses earlier feature maps and supports gradient flow through the network.
Dense connectivity
Earlier feature maps are reused
03 · Transfer learning · ImageNet
EfficientNetB0
Compound scaling balances model depth, width, and input resolution.
Scaling strategy
Considers depth, width, and resolution together for an accuracy-to-computation comparison.
04 · Transfer learning · ImageNet
ResNet50
Residual connections create shortcut paths through a deeper network.
Skip connections
Shortcut paths help information and gradients travel through deeper layers.
Common evaluation
Compare performance & efficiency
Quadratic Weighted KappaMacro F1ROC-AUCTraining time
Technology & Study Insights
The tools behind the study.
Frameworks, preprocessing & explainability
- CNN
- CLAHE
- Grad-CAM
Study spotlight · Transfer learning
DenseNet121Reuse features. Keep information flowing.
DenseNet connects each layer to the feature maps from earlier layers. This dense connectivity promotes feature reuse and helps gradient flow through the network. In this study, DenseNet121 uses ImageNet transfer learning and is compared with three other CNN approaches.
01
Dense connections
Layers receive earlier feature maps.
02
Feature reuse
Previously learned patterns stay available.
03
Transfer learning
ImageNet weights initialize the model.
Image preparation
Make retinal details clearer.
The workflow crops dark borders, resizes and normalizes images, and explores CLAHE / Gaussian blend contrast enhancement before training.
Explainable predictions
Look beyond the class label.
Grad-CAM maps help visualize which image regions influenced a model prediction; they support interpretation and are not clinical evidence.
From image to severity class
Fundus imageAPTOS 2019
PreprocessingCrop · resize · enhance
DenseNet121ImageNet transfer learning
Severity gradeOne of five classes
Experimental Workflow
From fundus image to model review.
The study uses the APTOS 2019 Blindness Detection training set: approximately 3,660 labeled RGB retinal images. Dataset files are not redistributed with the project.
01
Explore
Review class distribution, image dimensions, and representative samples.
02
Preprocess
Crop dark borders, resize, enhance contrast, and normalize fundus images.
03
Split
Create stratified train, validation, and held-out test sets.
04
Balance
Apply class weighting to address the natural skew toward No DR cases.
05
Train
Compare all four architectures with matched data splits and augmentation.
06
Explain
Use Grad-CAM to inspect which retinal regions influence predictions.
Evaluation & Explainability
Measure performance. Inspect attention.
Quadratic Weighted Kappa is the primary metric for the ordered grades. Accuracy, per-class and macro precision/recall/F1, one-vs-rest ROC-AUC, confusion matrices, parameter counts, and training time provide additional comparisons. Grad-CAM is used to visualize influential retinal regions.
ModelVal. κTest κAccuracy
Custom CNNPendingPendingPending
DenseNet121PendingPendingPending
EfficientNetB0PendingPendingPending
ResNet50PendingPendingPending
Numerical results were not included in the supplied project summary and are intentionally not estimated here.
Team Contributions
A collaborative model study.
Developed for the SE4050 Deep Learning module at SLIIT. Review the final notebooks in the Finalize-Models branch.
Browse final notebooksBagya R M S
Custom CNN · Preprocessing pipeline
02
Insath MMM
DenseNet121
03
Gayathree M.G.K
EfficientNetB0
04
Purijjala R W M A H
ResNet50 · Evaluation · Grad-CAM
Up Next in Portfolio · Final Year Research · Disaster Risk
Integrated Flood and Landslide Risk Assessment & Response System→
A Sri Lanka-focused research prototype connecting flood impact mapping, explainable landslide prediction, GN-level human impact assessment, and hazard-aware emergency response.
PythonMachine LearningGeoPandasFastAPIWeb GISOR-Tools