Speaker
Description
Recent advances in unsupervised segmentation, particularly with transformer-based models like MAESTER, have shown promise in segmenting Electron Microscopy (EM) data at the pixel level. However, despite their success, these models often struggle with capturing the full hierarchical and complex nature of EM data, where variability in texture and the intricate structure of biological components pose significant challenges. To address these limitations, I employ a hierarchical variational autoencoder (VAE), which I believe is better suited for this task due to its ability to naturally capture and represent the hierarchical structure inherent in EM images. This approach, enhanced with contrastive loss and sparse ground truth annotations, effectively structures the latent space, allowing for the clear separation of subcellular structures and improving segmentation accuracy. Additionally, I implement a masking strategy within an inpainting task, where the network predicts masked pixels, ensuring that the latent space robustly represents diverse EM structures. While still being optimized, this method has demonstrated promising progress, aligning with the I2K conference’s goals to transform images into actionable knowledge and offering substantial potential to advance biological research and understanding.
| Authors | Sheida R. Kordasiabi*, Florian Jug |
|---|---|
| Keywords | Electron microscopy, semantic segmentation, contrastive learning |
