Diabetic Retinopathy Detection M.Tech Project: EfficientNet-B0 with CBAM Attention, Ordinal Loss and Grad-CAM

About the diabetic retinopathy detection M.Tech project

Diabetes slowly damages the tiny blood vessels of the retina, and without early screening a patient can lose sight for good. This diabetic retinopathy detection M.Tech project grades a colour fundus photograph into five stages (No DR, Mild, Moderate, Severe, Proliferative) and draws a heatmap that shows which spots led to the grade.

The research question is practical: can one light network with attention match a heavy two-branch CNN and transformer model, respect the order of the grades, and still work on images from a different camera?

Research behind it

Base paper. The study starts from A hybrid neural network approach for classifying diabetic retinopathy subtypes, published in Frontiers in Medicine (2024). It joins an EfficientNet branch for local lesions with a Swin Transformer branch for global context, fuses both, and tests on the APTOS dataset against ResNet101, VGG19, InceptionV3, EfficientNet-B3 and Swin Transformer V2.

Research gap. Two full backbones are heavy for a CPU-only clinic computer. The five grades have a natural order, yet a plain classifier treats a Mild-to-Moderate mistake the same as a Mild-to-Proliferative one. Testing on one dataset says little about other cameras, and class activation maps are shown as pictures without a measured check.

Proposed system and novelty. A single EfficientNet-B0 gets a CBAM block that weighs channels and image regions, so small micro-aneurysms are not drowned out by the bright optic disc. An ordinal head predicts "grade above 0, above 1, above 2, above 3", which teaches the model the order of severity. The model is also tested on IDRiD fundus images without retraining, and its Grad-CAM heatmaps are scored against IDRiD lesion masks with a pointing-game check.

Models compared. ResNet50, plain EfficientNet-B0, Swin-Tiny and the base paper hybrid re-implemented on the same split, against the proposed model. Metrics are accuracy, macro F1, quadratic weighted kappa, parameters and CPU time. The ablation study removes CBAM, swaps the ordinal loss for cross-entropy and skips blur subtraction.

For a related imaging topic, see the chest X-ray pneumonia and TB detection project, which studies a hybrid CNN-ViT.

How it works

• The doctor creates a patient record and uploads the left and right eye images.
• A quality check rejects images that are too blurred or too dark.
• The image is cropped to the retina, enhanced by blur subtraction and resized.
• The CBAM model returns the grade with an ordinal confidence bar.
• Grad-CAM builds a heatmap laid over the original photo.
• The doctor adds notes and referral advice and downloads a PDF report.
• The admin compares all trained models, including the base hybrid, on one page.

Users and modules

There are three roles. The doctor screens patients, reviews heatmaps, writes notes and prints reports. The patient logs in to see past reports and follow-up reminders. The admin manages users and records, checks the dashboard and the model comparison page, and sets the referral grade. Training runs on Google Colab, while the trained model grades images on a normal laptop CPU.

What you get

• Phase 1 Research: Base paper study: EfficientNet and Swin Transformer hybrid (Frontiers in Medicine, 2024); Literature survey of 15-20 recent DR grading papers; Research gap, novelty and objectives; APTOS 2019 and IDRiD EDA notebook; Synopsis and Phase-1 report.
• Phase 2 Implementation: Flask web app with doctor, patient and admin panels; Preprocessing pipeline (crop, blur subtraction, resize); Baselines and base hybrid re-implementation; Proposed EfficientNet-B0 + CBAM + ordinal head; Grad-CAM module with pointing-game check; MySQL database and trained models.
• Results: Accuracy, macro F1 and quadratic kappa; External test on IDRiD; Ablation study graphs and CPU time table; results.xlsx and all graphs in high resolution.
• Thesis and paper: Thesis written for you (about 90 pages); IEEE conference paper draft; Journal paper draft; Publication guide.
• Reviews, viva and support: PPTs for Review 1, Review 2 and final viva; Viva book with 50+ questions; A1 poster and project diary; Demo and code walkthrough videos; 1-on-1 setup, 2 revisions and support till viva.

Who can use this project

M.Tech and M.E. students of CSE and IT who need a medical imaging dissertation with a clear base paper will find this a good fit. It also suits searches for diabetic retinopathy detection using deep learning, an EfficientNet CBAM attention project or an ordinal regression medical imaging project. The Keras code is short and commented, so each layer can be explained in the viva.

Frequently asked questions

Which base paper and fundus datasets does this grading work use?

The base paper is the EfficientNet and Swin Transformer hybrid published in Frontiers in Medicine (2024). Training uses the public APTOS 2019 dataset, and IDRiD is used only as an unseen test set.

Why use an ordinal loss instead of normal classification?

Grades follow an order, so predicting Severe for a Mild eye is a worse error than predicting Moderate. The ordinal head and quadratic kappa reward predictions that stay close to the true stage.

Do I need a GPU to run the grading app?

No. The trained model runs on a CPU laptop. Full training of all five models is done on Google Colab with the notebooks included.

Will my thesis match another student's thesis?

No. Every buyer gets a separate title variant, a different model change and a thesis written fresh. Before delivery the text is compared with every earlier copy of this project and must stay under 10% match.

How is the Grad-CAM heatmap checked for quality?

IDRiD provides expert masks of lesions. A pointing-game score counts how often the hottest heatmap spot falls inside a marked lesion, which gives the results chapter a measured answer instead of a picture alone.