Boston University
A survey of diffusion models and their applicabillity to digital staining of histological specimens
Abstract
dc:description.abstractIn the quest to deepen our understanding of the human brain, neuroscience increasingly depends on sophisticated imaging techniques capable of revealing the organ’s complex structures at multiple scales. While traditional histological methods like Gallyas silver myelin staining offer valuable insights into micro- and meso-scale features, they struggle with issues of staining variability, tissue damage, and distortion, complicating accurate 3D reconstruction. The advent of digital staining (DS) frameworks using CycleGANs (Zhu et al., 2020), Contrastive Unpaired Translation (CUT) (Park et al., 2020), and serial sectioning optical coherence tomography (S-OCT) (Min et al., 2020) has marked significant progress, albeit with limitations such as GAN-based models’ training instability. Addressing these challenges, we survey the generative diffusion models, despite their inherent challenges related to extensive data requirements and lengthy training times. We specifically assess the suitability of these models for digital staining of histological samples. Integrating the CLIP (Contrastive Language-Image Pretraining)(Radford et al., 2021) model, renowned for its semantic richness, we adapt its capabilities to meet the high-resolution demands typical of histology. Preliminary results from our training experiments demonstrate this approach’s potential. Furthermore, we introduce a novel methodology named CLIPpers (CLIP with Pretrained Contrastive Patch-Image Pairs), which leverages semantic embeddings from CLIP to enhance diffusion model training in biomedical imaging. This approach addresses the challenges of data scarcity and training inefficiency by utilizing the robust feature extraction capabilities of CLIPs vision encoder. CLIPpers employs global semantic conditions to direct the diffusion process, ensuring high image fidelity and accelerated training. Additionally, the integration of a contrastive loss mechanism refines the models proficiency in handling high-resolution images and performing patchwise comparisons. This optimization significantly enhances the models ability to accurately discern and replicate relevant features. Our findings suggest that CLIPpers could have a profound impact on biomedical research and applications, offering a robust and efficient solution to the longstanding challenges of resolution enhancement and data scarcity in biomedical imaging.
Degree
thesis:*- Name thesis:degree_name
- Master of Science
- Level thesis:degree_level
- masters
- Discipline thesis:degree_discipline
- Computer Science
- Grantor
- Boston University
- Year dc:date.issued
- 2024
Author and committee
dc:creator, dc:contributor.*- Author dc:creator
-
- Wu, Yicun
- Advisor dc:contributor.advisor
-
- Gardos, Thomas
Subjects
dc:subject × 4Rights
- Language dc:language.iso
- en_US
Identifiers
dc:identifier.*- Handle dc:identifier.uri
- https://hdl.handle.net/2144/53513