Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

CHAMMI-75: pre-training multi-channel models with heterogeneous microscopy images

Vidit Agrawal, John Peters, Tyler N. Thompson, Mohammad Vali Sanian, Chau Pham, Nikita Moshkov, Arshad Kazi, Aditya Pillai, Jack Freeman, Byunguk Kang, Samouil L. Farhi, Ernest Fraenkel, Ron Stewart, Lassi Paavolainen, Bryan A. Plummer, Juan C. Caicedo

TL;DR

CHAMMI-75 introduces a large, heterogeneous multic-channel microscopy dataset designed to train channel-adaptive and channel-agnostic models for cell morphology. By pre-training ViT-based architectures with self-supervised learning on CHAMMI-75, the authors demonstrate strong cross-benchmark generalization, especially for novel channel configurations and modalities, with bag-of-channels often outperforming multi-channel attention in SSL settings. The work shows that dataset diversity and modality breadth are key drivers of robust representations and that these representations can be disentangled from technical batch effects via standard batch-correction techniques. The dataset and accompanying benchmarks provide a foundation for scalable, cross-lab cellular morphology models with potential to generalize across imaging technologies and study designs.

Abstract

Quantifying cell morphology using images and machine learning has proven to be a powerful tool to study the response of cells to treatments. However, models used to quantify cellular morphology are typically trained with a single microscopy imaging type. This results in specialized models that cannot be reused across biological studies because the technical specifications do not match (e.g., different number of channels), or because the target experimental conditions are out of distribution. Here, we present CHAMMI-75, an open access dataset of heterogeneous, multi-channel microscopy images from 75 diverse biological studies. We curated this resource from publicly available sources to investigate cellular morphology models that are channel-adaptive and can process any microscopy image type. Our experiments show that training with CHAMMI-75 can improve performance in multi-channel bioimaging tasks primarily because of its high diversity in microscopy modalities. This work paves the way to create the next generation of cellular morphology models for biological studies.

CHAMMI-75: pre-training multi-channel models with heterogeneous microscopy images

TL;DR

CHAMMI-75 introduces a large, heterogeneous multic-channel microscopy dataset designed to train channel-adaptive and channel-agnostic models for cell morphology. By pre-training ViT-based architectures with self-supervised learning on CHAMMI-75, the authors demonstrate strong cross-benchmark generalization, especially for novel channel configurations and modalities, with bag-of-channels often outperforming multi-channel attention in SSL settings. The work shows that dataset diversity and modality breadth are key drivers of robust representations and that these representations can be disentangled from technical batch effects via standard batch-correction techniques. The dataset and accompanying benchmarks provide a foundation for scalable, cross-lab cellular morphology models with potential to generalize across imaging technologies and study designs.

Abstract

Quantifying cell morphology using images and machine learning has proven to be a powerful tool to study the response of cells to treatments. However, models used to quantify cellular morphology are typically trained with a single microscopy imaging type. This results in specialized models that cannot be reused across biological studies because the technical specifications do not match (e.g., different number of channels), or because the target experimental conditions are out of distribution. Here, we present CHAMMI-75, an open access dataset of heterogeneous, multi-channel microscopy images from 75 diverse biological studies. We curated this resource from publicly available sources to investigate cellular morphology models that are channel-adaptive and can process any microscopy image type. Our experiments show that training with CHAMMI-75 can improve performance in multi-channel bioimaging tasks primarily because of its high diversity in microscopy modalities. This work paves the way to create the next generation of cellular morphology models for biological studies.
Paper Structure (64 sections, 23 figures, 24 tables)

This paper contains 64 sections, 23 figures, 24 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. The CHAMMI-75 Dataset
  4. Evaluation benchmarks
  5. Experiments and Results
  6. Benchmarking experiments
  7. Channel Generalization and Cross Modality Transfer
  8. Factors impacting representation learning
  9. Scalability analysis
  10. Weakly supervised learning
  11. Additional Analysis
  12. Conclusion and Limitations
  13. Dataset
  14. Data Preparation
  15. Dataset Curation
  16. ...and 49 more sections

Figures (23)

  • Figure 1: Example images in CHAMMI-75, a heterogenous dataset of multi-channel microscopy images. Channel colors are assigned to an RGB pseudo-color for visualization.
  • Figure 2: Comparison of existing microscopy datasets used for representation learning of cell morphology. CHAMMI-75 is the largest dataset of multi-channel microscopy images. Others datasets: IDRCell100k bourriez2024chada, CHAMMI chen2023chammi, CytoImageNet hua2021cytoimagenet, HPAv23 gupta2024subcell, Microsnoop xun2024microsnoop, RxRx rxrx, JUMP-CP chandrasekaran2023jump, and Phenoprints-16M kenyon2024vitally.
  • Figure 3: Diversity of sources and biological studies in CHAMMI-75. The treemap illustrates the distribution of images according to the hosting platforms they were obtained from (colors) and the type of biological study (inner rectangles). We sampled from 18 different sources to ensure broad coverage of biological study types.
  • Figure 4: Content and distribution of images in CHAMMI-75 according to the integrated metadata. a) Selected metadata fields and summary statistics of their diversity. b) Sparse, long-tail distribution of channel configurations across studies. None of the studies has all channel types, and none of the channels is used in all studies.
  • Figure 5: Dataset curation pipeline. Left: the dataset downloaded from the hosting platforms. Middle: the metadata is used to filter redundancy by randomly sampling a few: 2D slices from 3D images, frames from live microscopy videos, and wells from control conditions. Right: content-based clustering selects diverse, high-quality images as a final step to create CHAMMI-75.
  • ...and 18 more figures