A team of undergraduate students from UNIST placed third at the 2026 AI in Healthcare Hackathon,  an international competition focused on medical image analysis. 

Competing as Team ULSANER, the four-member group advanced to the final round held in Uzbekistan from March 27–28, 2026. Among 58 finalist teams from 17 countries, the team achieved the top score in the initial model evaluation stage before securing third place overall following final presentations and technical assessments. 

Organized by Central Asian University as part of the CAU Tech Hackathon 2026, the competition drew more than 1,200 participants across approximately 300 teams worldwide. 

Team ULSANER, consisting of Jeong Jae Lee and MinSeong Kim from the Department of Industrial Engineering, Minguk Jeon from the Department of Computer Science and Engineering, and HyeRi Cho from the Department of Biomedical Engineering. Their interdisciplinary collaboration enabled the integration of computational methods with domain-specific insights. 

The final challenge required participants to develop an AI system, capable of classifying 12 types of skin lesions from biopsy images and identifying lesion regions through segmentation. Evaluation criteria extended beyond model accuracy to include clarity of presentation and practical usability.

To address this task, the team developed a deep learning–based framework integrating classification and segmentation. Using an automated optimization approach, they systematically evaluated model architectures and hyperparameters before applying an ensemble method to ensure consistent performance across new data. 

Model robustness was strengthened through cross-validation, focal loss to address class imbalance, and data augmentation techniques. This approach improved segmentation performance, increasing the Intersection over Union (IoU) metric from 0.8474 to 0.8528. 

To demonstrate real-world applicability, the team implemented a web-based interface, SkinScanner, enabling users to upload images and view both classification results and segmented outputs. The interface incorporated the clinically recognized ABCD rule—asymmetry, border, color, and diameter—to enhance interpretability and user understanding. 

“The project brought together our academic training and collaborative problem-solving across disciplines," said Jeong Jae Lee, who led the team. “Refining the system through each stage of the competition was particularly meaningful.” 

MinSeong Kim noted that anomaly detection in complex data has broad applications beyond healthcare, while Minguk Jeon emphasized the efficiency gained through automated optimization. HyeRi Cho added that the experience highlighted the importance of interpretability and trust in AI systems.