Multi-tile image stitching aims to merge multiple natural or biomedical images into a single mosaic. This is an essential step in whole-slide imaging and large-scale pathological imaging systems. To tackle this task, a multi-step framework is usually used by first estimating the optimal transformation for each image and then fusing them into a whole image. However, the traditional approaches are usually time-consuming and require manual adjustments. Advances in deep learning techniques provide an end-to-end solution to register and fuse information of multiple tile images. In this paper, we present a deep learning model for multi-tile biomedical image stitching, namely MosaicNet, consisting of an aligning network and a fusion network. We trained the MosaicNet network on a large simulation dataset based on the VOC2012 dataset and evaluated the model on multiple types of datasets, including simulated natural images, mouse brain T2-weighted Magnetic Resonance Imaging (T2w-MRI) data, and mouse brain polarization sensitive-optical coherence tomography (PS-OCT) data. Our method outperformed traditional approaches on both natural images and brain imaging data. The proposed method is robust to different settings of hyper-parameters and shows high computational efficiency, up to approximately 32 times faster than the conventional methods.