DSnet: Removing scattered light in biomedical images using a Dual Stream neural network

Light scattering is a common phenomenon in biomedical imaging. However, its removal is a challenging task because of the complex light paths from the scattered light. In this paper, we propose a dual-stream neural network to computationally remove the scattered light in biomedical images. To achieve this task, we first derive a mathematical model from the classical dehazing model for natural images. Then, we propose a dual-stream neural network to solve this model. Different from the dehazing task for natural images that have clear images as ground truth, biomedical images usually do not have the ground truth. Therefore, we use images from two different modalities to train the dual-stream network without knowing the ground truth. Several experiments confirm that our method can significantly reduce the light scattering and improve the image quality. The proposed method can be applied in various applications where removing the scattered light is needed or preferred.