SliceAutoMapi: High-throughput Automated Mapping of 2D Mouse Brain Image Slices to 3D Brain Volumes

High-throughput spatial transcriptomics technologies such as MERFISH and Stereo-Seq generate massive brain sections requiring accurate registration to 3D reference atlases. However, existing methods face critical limitations. Manual tools lack scalability, feature-based methods suffer from anatomical landmark loss due to technical variability, and existing deep learning approaches are constrained by annotation scarcity and input size limits. Herein, we propose a ResNet-Transformer slice-to-volume registration method for accurate, high-throughput alignment of brain sections across coronal, sagittal, and horizontal orientations (SliceAutoMapi). Our approach leverages isotropic datasets to systematically generate multi-orientation training images, integrates ResNet for global feature extraction with Transformer for local and inter-slice spatial learning, and incorporates priors including slice spacing to improve slices registration efficiency. Experimental results demonstrate that our method achieves lower voxel and angle errors and a higher Dice coefficient on the controlled test set, validating its advantages for spatial transcriptomics applications. In conclusion, the hybrid ResNet- Transformer registration framework effectively addresses the core limitations of slice registration methods. It enables robust, high-throughput 3D atlas alignment for multi-plane brain sections and provides a reliable registration solution for large-scale brain tissue analysis.