The exorbitant cost of accurately annotating the large-scale serial scanning electron microscope (SEM) images as the ground truth for training has always been a great challenge for brain map reconstruction by deep learning methods in neural connectome studies. The representation ability of the model is strongly correlated with the number of such high-quality labels. Recently, the masked autoencoder (MAE) has been shown to effectively pre-train Vision Transformers (ViT) to improve their representational capabilities.
In this paper, we investigated a self-pre-training paradigm for serial SEM images with MAE to implement downstream segmentation tasks. We randomly masked voxels in three-dimensional brain image patches and trained an autoencoder to reconstruct the neuronal structures.
We tested different pre-training and fine-tuning configurations on three different serial SEM datasets of mouse brains, including two public ones, SNEMI3D and MitoEM-R, and one acquired in our lab. A series of masking ratios were examined and the optimal ratio for pre-training efficiency was spotted for 3D segmentation. The MAE pre-training strategy significantly outperformed the supervised learning from scratch. Our work shows that the general framework of can be a unified approach for effective learning of the representation of heterogeneous neural structural features in serial SEM images to greatly facilitate brain connectome reconstruction.