DISSECT: Diffusion-based Cell Segmentation Integrating Spatial Transcriptomics

DISSECT (DIffusion-based cell Segmentation integrating Spatially rEsolved transCripTome) is an advanced cell segmentation algorithm developed by the Zeng Lab. It combines cytological image segmentation with transcriptome-guided fine-tuning to achieve precise spatial transcriptome reconstruction.

DISSECT leverages a pre-trained deep generative model to identify cell nuclei or membrane boundaries. By unifying the gradient fields of cytological images and transcriptomics data, it refines cell boundaries and reconstructs spatial single-cell transcriptomes.

Key Features

• Integration of deep generative models for segmentation.
• Combines spatial transcriptomic data with image analysis for enhanced accuracy.
• Supports transcriptomics data from multiple platforms:
  • 10X Xenium
  • BGI Stereo-seq
  • Nanostring CosMx
• Outputs include refined cell masks, bounding boxes, and spatial transcriptomes for downstream analyses.

Dependencies

Below are the essential libraries required to use DISSECT:

• Python Packages:
  • numpy
  • pandas
  • torch
  • tqdm
  • scipy
  • skimage
  • shapely
  • detectron2
  • joblib

For full dependencies, refer to the environment.yml file.

Installation

Currently, DISSECT supports the following installation methods:

• Pypi: Coming soon.
• Conda: Coming soon.

Input Requirements

To use DISSECT, you need:

1. Aligned Cytological Images:

   • Supported formats: .tif, .tiff, .png, .jpg.
   • Images should be pre-registered with spatial transcriptomics data.
   • Recommended libraries for loading images: tifffile, cv2, or skimage.

2. Transcriptomics Data:

   • For platforms like 10X Xenium, BGI Stereo-seq, or Nanostring CosMx, provide the native gene matrix file.
   • For other platforms, prepare a .csv or .tsv file with the following four columns:
     • x: The x-coordinate of the transcript.
     • y: The y-coordinate of the transcript.
     • geneID: The name or identifier of the transcript.
     • MIDCount: The number of transcripts at the given location (usually 1 for imaging-based technologies).

Usage Example

import DISSECT

mask = DISSECT.segmentation(
    img_path="example_stereseq.tif",
    platform="stereoseq",
    gene_mtx_filename="example_stereoseq.gem",
    config_file="config.yaml",
    weights_file="model_weights.pth"
)

Parameters:

• img_path: Path to the cytological image file.
• gene_mtx_filename: Path to the gene matrix file in .csv, .tsv, or .gem format.
• platform: The sequencing platform, options include:
  • "xenium"
  • "stereoseq"
  • "nanostring"
• config_file: Path to an optional YAML file for model configuration parameters.
• weights_file: Path to the pre-trained model weights file (download available https://drive.google.com/file/d/1Y9_YCJzhUPEQBDAdKVyrKplI1vpD4qiO/view?usp=sharing).

Optional Parameters:

• output: Path to store the .npy file of segmented mask and detected bounding boxes. Default: Current working directory.
• alpha (float): Sets the weight of the transcriptomic gradient field in refining cell boundaries. Default: 0.5.
• expand_by (int): Number of pixels by which the bounding box is expanded outward. This ensures accurate cell boundaries without enlarging the actual cell size. Default: 5.
• gene (bool): Whether to use transcriptomics data for fine-tuning. Default: True.
• random_seed (int): Seed for random number generation to ensure reproducibility. Default: No predefined seed.
• num_proposals (int): Number of bounding boxes proposed by the diffusion model. Default: 1200.
• isslice (bool): Whether to extract a subset of the image for training. Default: False.
• x0, y0 (int): The x and y coordinates of the upper-left corner of the extracted region when isslice is True.
• length, width (int): The length (x-axis) and width (y-axis) of the extracted image region when isslice is True.
• fov (int): Field of view ID to select a specific local area in the gene expression matrix, applicable when platform="nanostring".

Outputs:

• Cell Masks: The model generates refined cell masks and saves them as a NumPy .npy file.
• Bounding Boxes: Detected cell bounding boxes are saved in a .csv file, allowing for further custom parameter adjustments.
• Spatial Transcriptomes: Outputs reconstructed spatial single-cell transcriptomes.

The output files are saved in the specified directory for downstream analysis.

Citation

If you use DISSECT in your research, please consider citing the following:

Zeng Lab
Dissection of Spatial Transcriptomics with Cytological Imaging Integration