Title

Revisiting Masked Image Modeling with Standardized Color Space for Domain Generalized Fundus Photography Classification

DGIST Authors

Eojin Jang ; Sang Hyun Park ; Sunghoon Im

Advisor

박상현

Co-Advisor(s)

Sunghoon Im

Issued Date

2025

Awarded Date

2025-02-01

Citation

Eojin Jang. (2025). Revisiting Masked Image Modeling with Standardized Color Space for Domain Generalized Fundus Photography Classification. doi: 10.22677/THESIS.200000828469

Type

Thesis

Description

Masked image modeling, Domain generalization, Fundus photography, Diabetic retinopathy

Table Of Contents

I. INTRODUCTION 1
II. RELATED WORKS 4
2.1 CFP Classification 4
2.2 Foundation Models 4
2.3 Domain Generalization (DG) 4
III. METHOD 6
3.1 Problem Setup 6
3.2 Overview 6
3.3 Masked Image Transformation 6
3.3.1 Standardized color transformation 6
3.3.2 Masked image modeling 8
3.4 Joint Representation Learning 9
3.4.1 Cross-Attention module 9
3.4.2 Classification with joint representation 10
IV. EXPERIMENTS 11
4.1 Datasets 11
4.1.1 4DR 11
4.1.2 APTOS-Messidor 11
4.1.3 Glaucoma 11
4.2 Experimental Details 11
4.3 Experimental Scenarios 12
4.3.1 Comparison against recent DG methods 12
4.3.2 Scalability in diverse datasets 12
4.3.3 Ablation studies 12
4.3.4 Impact on masked image modeling 13
4.3.5 Impact on transformed image 13
4.3.6 Impact on cross-attention 13
V. RESULTS 14
5.1 Comparison Against Recent DG Methods 14
5.2 Scalability in Diverse Datasets 14
5.3 Ablation Studies 14
5.4 Impact on Masked Image Modeling 15
5.5 Impact on Transformed Image 16
5.6 Impact on Cross-Attention 17
5.7 Additional Results 17
5.7.1 Impact on the blended image Zi 17
5.7.2 Impact of class label use on color transformation 18
5.7.3 Additional transformed images 20
VI. CONCLUSION 23
References 24

URI

http://hdl.handle.net/20.500.11750/58105
http://dgist.dcollection.net/common/orgView/200000828469

DOI

10.22677/THESIS.200000828469

Degree

Master

Department

Artificial Intelligence Major

Publisher

DGIST

Show Full Item Record