SplineDist: Automated Cell Segmentation with Spline Curves

This repository contains the implementation of SplineDist, a machine learning framework for automated cell segmentation with spline curves. The manuscript is accepted at ISBI 2021. The code in its current state allows reproducing the paper experiments but is still in development. We are currently working to package it in a cleaned and optimized form. In the meantime, we encourage interested end-users to contact us for more information and assistance.

Overview

SplineDist has been designed for performing instance segmentation in bioimages. Our method has been built by extending the popular StarDist framework. Our repository relies on the high-quality StarDist repository. We encourage the user to explore StarDist repository for further details on the StarDist method.

While StarDist models objects with star-convex polygonal representation, SplineDist models objects as parametric spline curves. As our representation is more general, it allows to model non-star-convex objects as well, with the possibility of conducting further statistical shape analysis.

Requirements

Tensorflow
StarDist 0.6.2 (can be installed with pip : pip install stardist==0.6.2)

Example installation instructions

Tested on Ubuntu 20.04. System compiler libraries required (installation of aptitude build-essential and gcc_linux-64 libraries may resolve compilation issues).

Initiate an environment titled splinedist containing python 3.6, tensorflow 2.4.0 and compilers required for build
```
 conda create -n splinedist -c conda-forge python=3.6 tensorflow=2.4.0 gcc_linux-64 gxx_linux-64 opencv -y
```
Additionally, for use of Jupyter notebooks:
```
 conda install ipykernel -y
```

Activate the new environment and install splinedist (from GitHub)

 conda activate splinedist
 git clone https://github.com/uhlmanngroup/splinedist.git
 cd splinedist
 pip install .

Walkthrough

Three walkthrough notebooks have been included in this repository for data-exploration, training, and inference tasks.

Datasets

The synthetic dataset used in the SplineDist manuscript can be found here.This dataset contains synthetic images with mostly star-convex and some non-star convex cell-like objects.