• • EllipseNet Descriptions
  • EllipseFit-MindSpore Descriptions
• Content
• • Model Architecture
  • Dataset
  • Environment Requirements
  • Quick Start
  • Script Description
  • Script Parameters
  • Training Process
  • Evaluation Process
  • • Evaluation
  • Model Description
  • • Performance
    • • Training Performance
      • Inference Performance
  • Citation
  • Contact

EllipseNet Descriptions

Paper: Chen, J., Zhang, Y., Wang, J., Zhou, X., He, Y. and Zhang, T., 2021, September. EllipseNet: Anchor-Free Ellipse Detection for Automatic Cardiac Biometrics in Fetal Echocardiography. In International Conference on Medical Image Computing and Computer-Assisted Intervention (pp. 218-227). Springer, Cham.

EllipseFit-MindSpore Descriptions

Please be noted that this is not a MindSpore version of EllipseNet, but using a 2D Unet to train a segmentation network and then using ellipses to fit the segmentation results. The details of this method can be found in ref[6] as compared in our paper. The PyTorch version of this algorithm can be found in https://git.openi.org.cn/OpenMedIA/EllipseNet

Content

• Content
  • EllipseNet Descriptions
  • Model Architecture
  • Dataset
  • Environment Requirements
  • Quick Start
  • Script Description
  • Script Parameters
  • Training Process
  • Evaluation Process
    • Evaluation
  • Model Description
    • Performance
      • Training Performance
      • Inference Performance
  • Citation
  • Contact

Model Architecture

The EllipseNet segementation network takes n 2D volumetric images as input, applies input and feature transformations. BN is introdued before each ReLU.

Dataset

Fetal Four Chamber View Ultrasound Scans

• Train: 225M, 1669 images with corresponding annotations.
• Test: 58M, 417 images with corresponding annotations.

Environment Requirements

• Hardware（Ascend）
  • Prepare hardware environment with Ascend processor.
• Framework
  • MindSpore
• For more information, please check the resources below：
  • MindSpore Tutorials
  • MindSpore Python API

• suwen package

  pip install -r requirements.txt
  pip install ./suwen-1.0.1-py3-none-any.whl
  

Quick Start

After installing MindSpore via the official website, you can start training and evaluation as follows:

# enter script dir, train PointNet
sh run_train_ascend.sh

# enter script dir, evaluate PointNet
sh run_eval.sh

Script Description

.
├── README.md
├── ellipsenet.ckpt
├── eval.py
├── eval_log.txt
├── requirements.txt
├── scripts
│   ├── run_eval.sh
│   ├── run_train_ascend.sh
│   └── scripts
│       ├── run_eval.sh
│       └── run_train_ascend.sh
├── src
│   ├── __init__.py
│   ├── __pycache__
│   │   ├── cross_entropy_with_logits.cpython-37.pyc
│   │   ├── dataset.cpython-37.pyc
│   │   ├── dice_loss.cpython-37.pyc
│   │   └── dice_metric.cpython-37.pyc
│   ├── config.py
│   ├── convert_nifti.py
│   ├── cross_entropy_with_logits.py
│   ├── dataset.py
│   ├── dice_loss.py
│   ├── dice_metric.py
│   ├── loss.py
│   ├── lr_schedule.py
│   ├── transform.py
│   └── utils.py
├── suwen-1.0.1-py3-none-any.whl
├── train.py
└── train_log.txt

4 directories, 27 files

Script Parameters

Major parameters in train.py are as follows:

--data_path: The absolute full path to the train and evaluation datasets.
--ckpt_path: The absolute full path to the checkpoint file saved after training.

More hyperparamteters can be modified in src/config.py.

Training Process

• running on Ascend

  sh run_train_ascend.sh
  

  After training, the loss value will be achieved as what in train_log.txt

  The model checkpoint will be saved in the current ckpt directory.

Evaluation Process

Evaluation

Before running the command below, please check the checkpoint path used for evaluation.

• running on Ascend

  sh scripts/run_eval.sh
  

  You can view the results through the file "eval_log". The accuracy of the test dataset will be as what in eval_log.txt.

Model Description

Performance

Training Performance

Parameters
Resource	Ascend 910; CPU 2.60GHz, 24cores; Memory, 96G
uploaded Date	11/29/2021 (month/day/year)
MindSpore Version	1.3.0
Dataset	Fetal Four Chamber View
Training Parameters	epoch=100
Optimizer	Adam
Loss Function	Softmax Cross Entropy
outputs	probability
Loss	Dice Loss
Checkpoint for Fine tuning	23M (ckpt file)

Inference Performance

Parameters
Resource	Ascend 910; CPU 2.60GHz, 24cores; Memory, 96G
uploaded Date	11/29/2021 (month/day/year)
MindSpore Version	1.3.0
Dataset	Fetal Four Chamber View
batch_size	1
outputs	probability
Dice	91.72%

Citation

If you find this project useful for your research, please cit our work use the following BibTeX entry.

@inproceedings{chen2021ellipsenet,
title={Ellipsenet: Anchor-free ellipse detection for automatic cardiac biometrics in fetal echocardiography},
author={Chen, Jiancong and Zhang, Yingying and Wang, Jingyi and Zhou, Xiaoxue and He, Yihua and Zhang, Tong},
booktitle={International Conference on Medical Image Computing and Computer-Assisted Intervention},
pages={218--227},
year={2021},
organization={Springer}
}

Contact

If you have any questions about this paper, welcome to email to zhangt02@pcl.ac.cn