Abstract
Purpose :
Retinal vessels analysis is crucial for early identification, diagnosis, and continuous monitoring of eye diseases. The extraction of quantitative metrics, such as width and tortuosity provides insights into a patient's overall health. The process of measuring vessels involves at first a necessary image processing step known as segmentation. The objective is to address limitations in current retinal vessels segmentation methods, particularly in the context of a scarcity of labeled images for training efficient Deep Learning algorithms.
Methods :
The proposed approach introduces four innovative methods specifically designed for arteries, veins, optic disk, and vessels as a whole. The segmentation process is streamlined, demanding only ten images of a specific dataset as input. This method aims to overcome the challenges posed by manual labeling's time-consuming nature and the inefficiency of current algorithms in scenarios with limited labeled images. Such scenarios are named Few-Shot learning and the Deep Learning algorithms we designed are dedicated to this purpose. The architecture of our four algorithms are derived from the well-known encoder-decoder U-Net, which cannot detect a single vessel in Few-Shot scenarios. We incorporate batch normalization to this architecture to solve two issues. First stabilizing the training with normalized data to avoid divergence of the learning process due to a lack of similarity between the data, which is a problem faced with few images. Then enhancing the prediction in real-world scenarios by avoiding using the learned normalization statistics. We made modifications to the depth of the architecture and the feature map space to further enhance details-gathering.
Results :
The method's efficacy is demonstrated with high superposition correspondence to manually labeled images: 69% for arteries, 73% for veins, 81% for vessels, and an outstanding 86% for the optic disk. These results showcase the breakthrough in addressing the limitations of existing techniques, especially in real-world conditions with a dearth of labeled images.
Conclusions :
In conclusion, the proposed method offers a significant advancement in retinal vessels segmentation. By reducing the dependency on a large number of manually labeled images and achieving remarkable superposition, this approach opens new possibilities for efficient and trustful retinal vessels analysis in diverse clinical scenarios.
This abstract was presented at the 2024 ARVO Imaging in the Eye Conference, held in Seattle, WA, May 4, 2024.