MedSight: A Web-Based Platform for Streamlined Medical Image Diagnosis and the Potential of Integrating Active Learning for Big Medical Data
1
Issued Date
2025-01-01
Resource Type
Scopus ID
2-s2.0-105017117940
Journal Title
2025 6th International Conference on Big Data Analytics and Practices Ibdap 2025
Start Page
58
End Page
63
Rights Holder(s)
SCOPUS
Bibliographic Citation
2025 6th International Conference on Big Data Analytics and Practices Ibdap 2025 (2025) , 58-63
Suggested Citation
Sinsawet N., Cheng S., Iamborisut P., Taurob S. MedSight: A Web-Based Platform for Streamlined Medical Image Diagnosis and the Potential of Integrating Active Learning for Big Medical Data. 2025 6th International Conference on Big Data Analytics and Practices Ibdap 2025 (2025) , 58-63. 63. doi:10.1109/IBDAP65587.2025.11145828 Retrieved from: https://repository.li.mahidol.ac.th/handle/123456789/112409
Title
MedSight: A Web-Based Platform for Streamlined Medical Image Diagnosis and the Potential of Integrating Active Learning for Big Medical Data
Author(s)
Author's Affiliation
Corresponding Author(s)
Other Contributor(s)
Abstract
In rural Thailand, many doctors face challenges in accessing diagnostic tools, resulting in the delay of treatments and the risk of misdiagnosis. To address this problem, this paper introduces MedSight, a web-based diagnostic support system that integrates project management, image annotation, model training, evaluation, and deployment functionalities. Though the current version has not included the active learning (AL) mechanism directly into the user interface, we have conducted separate investigations into AL's potential to reduce labeling costs while improving model performance. Our experiment compares the entropy-based AL with random sampling strategies in two medical imaging tasks: skin lesion segmentation and dental X-ray object detection. AL achieved a Dice score of 0.575, outperforming the random-selection baseline (0.376), using only 10% of the ISIC 2017 dataset. As for dental X-ray object detection, AL achieved a mAP@ 50 of 0.477 using 40% of the data, compared to 0.348 from the random-selection baseline. These results indicate that AL can accelerate model learning and enhance performance with fewer labeled examples by utilizing available large-scale, unlabeled medical data. MedSight offers solutions to the clinical requirements in low-resource settings and shows potential for future integration of AL strategies to improve diagnostic accuracy and reduce annotation burden.