Development of a Framework for Analyzing Lesion Classification Models Using Real and Generated Fundus Images 


Vol. 51,  No. 4, pp. 814-824, Apr.  2026
10.7840/kics.2026.51.4.814


PDF Full-Text
  Abstract

In the field of medical image analysis, deep learning models have shown high performance, but securing sufficient training data is limited by privacy concerns and data imbalance. This study proposes an evaluation framework to assess the performance and stability of six lesion classification models using datasets composed of various ratios of real and generated fundus images. The proposed framework computes confusion matrix– based metrics, regression, and statistical analysis indicators, and integrates them into a stability index for each model. The results show that models such as DenseNet121 and InceptionV3 maintained stable performance even with higher proportions of generated data, and in some cases, outperformed models trained only on real data. In contrast, MobileNetV2 and NASNetMobile exhibited sharp performance degradation and lower stability indices as the proportion of generated data increased. This study presents a systematic framework to evaluate and compare the effects of model architecture and data ratio when using generated data in medical image AI, and it can serve as a guideline for designing models and constructing datasets based on generated data.

  Statistics
Cumulative Counts from November, 2022
Multiple requests among the same browser session are counted as one view. If you mouse over a chart, the values of data points will be shown.


  Related Articles
  Cite this article

[IEEE Style]

H. Song and K. Song, "Development of a Framework for Analyzing Lesion Classification Models Using Real and Generated Fundus Images," The Journal of Korean Institute of Communications and Information Sciences, vol. 51, no. 4, pp. 814-824, 2026. DOI: 10.7840/kics.2026.51.4.814.

[ACM Style]

Ho-Jung Song and Ki-won Song. 2026. Development of a Framework for Analyzing Lesion Classification Models Using Real and Generated Fundus Images. The Journal of Korean Institute of Communications and Information Sciences, 51, 4, (2026), 814-824. DOI: 10.7840/kics.2026.51.4.814.

[KICS Style]

Ho-Jung Song and Ki-won Song, "Development of a Framework for Analyzing Lesion Classification Models Using Real and Generated Fundus Images," The Journal of Korean Institute of Communications and Information Sciences, vol. 51, no. 4, pp. 814-824, 4. 2026. (https://doi.org/10.7840/kics.2026.51.4.814)
Vol. 51, No. 4 Index

Search
(IN TITLE, AUTHOR, ABSTRACT,KEYWORDS)

Advanced Search ▼

POPULAR KEYWORDS
(TOP 10 KEYWORDS)

Recent Publications
(LAST 3 YEARS)






pISSN : 1226 - 4717
eISSN : 2287 - 3880