HiS4MAE: High-efficiency Segmentation of Subcellular Structure via Self-distillated Masked Autoencoder

Qizhe Liu, Huangxin Xu, Jue HOU, Jintu Zheng, Zenan Wang

Keywords: Subcellular Structure Segmentation, Masked Image Model, Microscopy Image

Abstract: The accurate identification of subcellular structures is crucial for understanding cellular functions. However, due to the varied morphology of different cells, conventional segmentation methods typically depend on a substantial collection of accurately labeled images of cell structures. The creation of such precise labels is often time-consuming and labor-intensive. To address this issue, we introduce an efficient, self-supervised method for segmenting subcellular structures, named HiS4MAE (High-efficiency Segmentation of Subcellular Structure via Self-distillated Masked Autoencoder). Leveraging an enhanced masked autoencoder (MAE), we train the encoder using the masked image modeling (MIM) framework, followed by clustering the encoded high-dimensional features to achieve pixel-level segmentation of structures. We employ a self-distillation technique to accelerate the model's training process and propose an inference method that is less time-consuming. We also introduce a discrete codebook to assist the self-distillation process, enhancing the model's stability during training. When applied to a publicly available volumetric electron microscopy (VEM) dataset of primary mouse pancreatic islet $\beta$ cells, HiS4MAE not only surpasses the state-of-the-art technique but also significantly reduces the time required for both training and inference.

Primary Area: applications to physical sciences (physics, chemistry, biology, etc.)

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Submission Number: 3695