Diabetic Retinopathy Grading using EfficientNetB0 (APTOS 2019 Dataset)

Diabetic retinopathy is one of the leading causes of preventable blindness, and it often progresses silently — by the time vision blurs, the damage can be severe. The standard defence is regular retinal screening: a clinician photographs the retina with a fundus camera and grades what they see on a 0-to-4 severity scale. The problem is scale and subjectivity. Grading thousands of screening photographs is slow, tiring work for scarce ophthalmologists, and early signs like a lone microaneurysm are easy for a tired human eye to miss. Automated grading does not replace the clinician, but it can triage: clearly normal images pass through quickly while suspicious ones are flagged for urgent expert review. This project implements exactly that triage pipeline — an EfficientNetB0 network fine-tuned on the clinician-labelled APTOS 2019 dataset, evaluated with the competition's own quadratic weighted kappa metric, and wrapped in a demo that shows the full journey from fundus photograph to graded report with per-grade explanations.

1. The 3,662 APTOS 2019 training images and their clinician-rated 0-4 labels are loaded and divided into stratified train and validation sets, preserving the grade distribution in both.
2. Every image is resized to 224x224 and normalised with ImageNet statistics; the training split additionally passes through augmentation (rotation, horizontal flip, zoom, brightness jitter) to simulate varied imaging conditions.
3. EfficientNetB0, initialised with ImageNet weights, is fine-tuned: early MBConv blocks stay frozen while later blocks and the new 5-way softmax head train with class-weighted categorical cross-entropy.
4. Training is monitored on the validation split using quadratic weighted kappa; the best checkpoint is retained and training stops before overfitting sets in.
5. The held-out validation set is evaluated exactly once: kappa, accuracy, confusion matrix and per-class precision/recall are generated and written into the report as the run's actual results.
6. At inference time, a new fundus photograph flows through the identical preprocessing and the saved model weights, and the web demo displays the 5-grade probability distribution with the predicted grade highlighted.

Tech stack:

• Python 3, TensorFlow/Keras
• EfficientNetB0 (ImageNet weights)
• OpenCV, Albumentations (augmentation)
• NumPy, pandas, scikit-learn (metrics)
• Jupyter Notebook (training & evaluation)
• APTOS 2019 Blindness Detection dataset
• HTML/CSS/JavaScript (grading web demo)
• Git

Dataset & model details

• Dataset: APTOS 2019 Blindness Detection — Kaggle competition by the Asia Pacific Tele-Ophthalmology Society; 3,662 clinician-labelled training fundus photographs; 5 severity classes verified against the official data description: 0 No DR, 1 Mild, 2 Moderate, 3 Severe, 4 Proliferative DR.
• Task: 5-class image classification; input = 224x224x3 fundus image, output = probability distribution over the five DR severity grades.
• Model: EfficientNetB0 transfer learning — ImageNet-pretrained MBConv backbone with compound scaling, global average pooling, and a dense 5-way softmax head (approximately 5.3M parameters, architecture design target).
• Metrics: Quadratic weighted kappa (the competition's official metric), accuracy, per-class precision/recall/F1, confusion matrix. No accuracy is claimed as measured — final performance depends on the executed training run (design target: approximately 0.80+ validation kappa, stated as a target, not a result).

• [EfficientNetB0 transfer learning] ImageNet-pretrained EfficientNetB0 is fine-tuned on APTOS 2019, with its original classifier replaced by a 5-way softmax head matching the five DR severity grades — full rationale documented for the viva.
• [Complete preprocessing pipeline] Fundus images are resized to 224x224, normalised with ImageNet statistics, and augmented during training (rotation, flips, zoom, brightness jitter) so the model tolerates varied cameras and lighting.
• [Quadratic weighted kappa evaluation] The APTOS competition's official metric, computed on the validation split alongside accuracy, per-class precision/recall and a confusion matrix — no cherry-picked numbers.
• [Class-imbalance handling] APTOS 2019 is heavily skewed toward No-DR images; stratified splits and class-weighted loss keep the rare severe and proliferative cases learnable.
• [Grading web demo] Upload any fundus photograph (or try the built-in samples) and receive the predicted grade with the full 5-class probability distribution and a plain-language explanation of each grade.
• [Grade reference guide] All five severity grades documented with their typical findings — microaneurysms, hemorrhages, hard exudates, venous beading, neovascularization — so the student can defend every prediction in the viva.
• [Referability flag] Grades 2-4 are flagged as referable DR, mirroring how real screening programs triage patients toward ophthalmologists; implemented as transparent demo logic, clearly labelled educational.

• Complete training & evaluation Jupyter notebook (preprocessing, augmentation, fine-tuning, metrics)
• Fine-tuned EfficientNetB0 weights with inference code
• Grading web demo: upload or sample images, probability distribution, per-grade explanations
• Confusion matrix, per-class metrics and kappa score from the actual training run
• Project report PDF (background, DR staging, methodology, results)
• PPT presentation for final review
• Viva Q&A preparation document (transfer learning, EfficientNet scaling, kappa, class imbalance, medical-AI limits)