Overview

This framework is provided to perform pixel-level segmentation of human liver, spleen, pancreas and kidney, based on MR images provided by German National Cohort(NAKO Dataset), using deep-learning method, and visualized the results. It establishes all functionality needed to operate on 3D images with a patch-based architecture.

NAKO Dataset:

  • Over 3400 labeled MRI images from thousands patients
  • Over 500 MRI images for evaluation

Used network architectures including 3d u-net, non-local neural network, attention u-net are proposed.

Arxiv:

Installation

use pip3 (with a venv)

pip3 install -e .

if it fails consider

pip3 install -e . --user

Usage

For training use

nohup python3 -u train.py > file_out 2> file_err &

For prediction use

nohup python3 -u evaluate.py > file_out 2> file_err &

Algorithm

non-local neural network

Inspired by the popular NLP Transformer architecture proposed by Google in 2017, an architecture of similar idea is proposed for image processing, the non-local neural networks.

It can capture the long-range dependencies between pixels more properly, check the paper from Wang Xiaolong https://arxiv.org/abs/1711.07971

Its architecture as following:

3d U-net as baseline

Baseline architecture is a 4-stages 3d u-net, as following:

Results

Achieve an average accurancy of 97% of all classes.