Seminar: Graduate Seminar
Machine learning for diabetic retinopathy detection from fundus images
Date:
March,21,2024
Start Time:
12:30 - 13:30
Location:
Auditorium 201, Biomedical Engineering
Add to:
Lecturer:
Yevgeniy Men
Research Areas:
| Diabetic retinopathy (DR) is a prevalent complication of diabetes associated with a significant risk of vision loss. Timely identification is critical to curb vision impairment. Algorithms for DR staging from digital fundus images (DFIs) have been recently proposed. However, models often fail to generalize due to distribution shifts between the source domain on which the model was trained and the target domain where it is deployed. A common and particularly challenging shift is often encountered when the source- and target-domain supports do not fully overlap. In this research, we introduce DRStageNet, a deep learning model designed to mitigate this challenge. We used seven publicly available datasets, comprising a total of 93,534 DFIs that cover a variety of patient demographics, ethnicities, geographic origins, and comorbidities. First, we developed a baseline classical machine learning model. A deep learning model is used to segment the arterioles, venules and lesions. Vasculature and lesion features are engineered and used as input to a gradient boosting model. Feature importance showed a strong reliance on DR lesion counts. However, the ablation study also showed that vasculature biomarkers contributed to the overall performance. This baseline model exhibited a fair performance in staging DR but still insufficient for translational research. As a second step we focused on the development of DRStageNet. We fine-tuned DINOv2, a pretrained model of self-supervised vision transformer, and implement a multi-source domain fine-tuning strategy to enhance generalization performance. DRStageNet significantly outperforms the baseline classical machine learning model as well as two state-of-the-art deep learning benchmarks, including a recently published foundation model. In particular, DRStageNet demonstrated to generalize well across the different datasets. We adapted the grad-rollout method to our regression task in order to provide high-resolution explainability heatmaps. The error analysis showed that 59% of the main errors had incorrect reference labels. We release the final DRStageNet model on our laboratory iOS platform, Lirot.ai, which is committed to making AI-driven analysis of ophthalmology images widely available, simplifying the integration of AI in the study of ophthalmology and promoting equitable access to AI technology in the field for research and educational purposes. M.Sc. student under the supervision of Prof. Joachim Behar.
|