Choroidal Thickness Estimation based on Fundus images with Uncertainty Quantification

Choroidal thickness estimation is a crucial step in monitoring the progression of choroid-related diseases. With the resurgence of deep learning technology, researchers begin to explore choroidal thickness estimation methods using fundus images based on convolutional neural networks (CNNs). Nonetheless, existing methods face challenges such as low accuracy and robustness. To improve the prediction accuracy and reliability of choroidal thickness estimation task using fundus images, a deep ensemble method with uncertainty quantification is proposed.. In general, the network architectures in ensemble methods are designed by hand, but hand-tuned network architectures which is time consuming and requires abundant experience. In this paper, we propose a novel choroidal thickness estimation method with network architecture search (NAS) guided by uncertainty quantification. By regulating the search direction, the proposed method aims to search the high accuracy and robustness models. Additionally, a selection method is proposed to select some of the searched models from all the searched models to make a balance between the performance in accuracy and robustness. Extensive experimental results show that the proposed method improves prediction accuracy and exhibits excellent uncertainty quantification performance for choroidal thickness estimation.