Skip to content

Latest commit

 

History

History
49 lines (34 loc) · 1.65 KB

README.org

File metadata and controls

49 lines (34 loc) · 1.65 KB
Raw

Self-supervised Non-Negative PU Learning + KSPTrack

Synopsis

This repository contains the code related to the method SSnnPU+KSPTrack from paper A Positive/Unlabeled Approach for the Segmentation of Medical Sequences using Point-Wise Supervision.

graphical_abstract.png

Installation

The following instructions have been tested to work on a Ubuntu 20.04 machine.

  • Clone this repository and head to the root directory.
  • Install PyTorch and pybind11 as build-time dependencies. This is necessary to build the CUDA superpixel pooling extension and the K-shortest paths C++ extension, respectively.
pip install "torch>=1.8" "pybind11>=2.6"
  • Install the package. This will install the required run-time dependencies.
pip install .

Datasets

We make the datasets, manual ground truth annotations and point-wise locations used in our experiments public. Head to https://zenodo.org/record/5007789/files/data.zip?download=1 to download the archive.

Usage

We provide a script to run the whole pipeline. Head to the ksptrack directory and run:

python run.py --in-path <path/to/DatasetXY> --out-path <path/to/results/DatasetXY> --pi-overspec-ratio 1.4 --cuda --ksp

The paths must be modified according to your use case. The ksp flag argument will run the spatio-temporal regularization step. The pi-overspec-ratio variable denotes a multiplicative factor applied to the maximum positive-class frequency computed from the manual ground truth annotations. It ultimately sets the initial prior (see variable eta in the paper for details).