The National AI Competition (NAIC) 2026 1st Runner-up

RetinaScope AI

(1) Introduction

The National AI Competition (NAIC) 2026 was organized by Sunway University and Rakan Tutor, culminating in the live Grand Finals at Sunway University on Saturday, June 13, 2026. I participated in this prestigious national hackathon with a team of 4. We joined the competition to push our boundaries in applied machine learning, test our collaborative engineering skills against real-world data constraints, and build a meaningful AI solution to address modern healthcare bottlenecks.

(2) The Task & Preparation

• The Task

Competing in the Computing Track under the theme Smart City Healthcare, our challenge was to build an automated 5-stage screening system for Diabetic Retinopathy (DR) using retinal fundus images, grading them from Stage 0 (No DR) to Stage 4 (Proliferative DR). The competition provided thousands of realistic images plagued by lighting variations, blurring, camera differences, and heavy class imbalance, and strictly prohibited external training data. The task demanded pure engineering pragmatism: maximizing the Quadratic Weighted Kappa (QWK) evaluation metric through smart preprocessing, architectural efficiency, fast inference, and lightweight deployment.

• Preparation

With a team of 4, we were able to divide and conquer the problem as a rigorous engineering challenge:-

Data & Preprocessing Strategy: We utilized Ben Graham preprocessing to normalize lighting variations across different fundus cameras and applied heavy data augmentations to make our networks robust against blur and artifacts.

Validation Strategy: To ensure our model generalized perfectly to unseen data without any data leakage, we engineered a strict Stratified Group K-Fold validation split loop.

Model Iteration: We spent weeks iterating from simple baselines to a sophisticated heterogeneous dual-stream architecture combining ConvNeXt-Tiny (for spatial hierarchies and global structures) and EfficientNet-B0 (for fine-grained textures like microaneurysms). We framed the problem as a regression task and calibrated classification boundaries using Nelder-Mead threshold optimization to squeeze out every bit of QWK performance.