Interpretable Disease Diagnosis Using Binocular Retinal Images with Vision Transformers
- 주제(키워드) Deep Learning , Multimodal Learning , Retinal Image , Disease Classification , Metabolic Syndrome
- 주제(DDC) 006.31
- 발행기관 아주대학교 일반대학원
- 지도교수 Kyung-Ah Sohn
- 발행년도 2025
- 학위수여년월 2025. 2
- 학위명 석사
- 학과 및 전공 일반대학원 인공지능학과
- 실제URI http://www.dcollection.net/handler/ajou/000000034283
- 본문언어 영어
- 저작권 아주대학교 논문은 저작권에 의해 보호받습니다.
초록/요약
Metabolic syndrome, marked by a combination of risk factors, substantially elevates the risk of diabetes, cardiovascular diseases, and other severe health issues. In this study, we utilized retinal images captured through fundoscopy during health check-ups to develop a predictive model for identifying metabolic syndrome. The model, trained exclusively on retinal images, achieved an AUC of 0.7752(±0.0034), demonstrating a promising ability to detect metabolic syndrome from retinal images. When we integrated basic clinical features such as age, gender, and body max index (BMI) with the retinal images, the model’s performance improved, achieving a significantly higher AUC of 0.8725(±0.0056). To further enhance the interpretability of our model and clarify the link between retinal features and metabolic syndrome, we apply a visualization method that highlights the specific retinal areas associated with metabolic syndrome. This visualization aids in understanding how different retinal regions contribute to the classification. Our findings underscore the potential of retinal images as a non-invasive tool for the early diagnosis of metabolic syndrome, opening avenues for improved preventive strategies.
more목차
1. Introduction 1
2. Related Works 5
2.1 Metabolic syndrome Prediction 5
2.2 Retinal Biomarkers in Systemic Disease 6
2.3 Deep Learning Applications in Retinal Image Analysis 7
2.4 Foundation Model 10
3. Methodology 12
3.1 Dataset 12
3.2 Feature Extraction and Concatenation 15
3.3 Explainable Visualization 16
4. Experiments 18
4.1 Experiment Details 18
4.2 Results 20
4.3 Interpretable Results 22
4.3.1 Feature Importance 22
4.3.2 Obesity-based Subgroup Analysis 23
4.3.3 Metabolic syndrome Associated Diseases 25
4.3.4 Explainable Visualization 27
4.4 Ablation Study 29
5. Conclusion 34
Discussion 36
Reference 39
