Skip to content

README.md

Latest commit

 

History

History
74 lines (59 loc) · 4.37 KB

README.md

File metadata and controls

74 lines (59 loc) · 4.37 KB

Multi-Disease, Multi-View & Multi-Center Right Ventricular Segmentation in Cardiac MRI (M&Ms2)

Caution

The pathology classes RV, TRI do not exist in training split, while exist in validation and test splits.

The M&Ms2 dataset contains 360 images (160 training, 40 validation and 160 testing) acquired from three clinical centers in Spain with 8 different scanners from 3 vendors in six hospitals across three countries. The pathologies include dilated left ventricle (DLV), dilated right ventricle (DRV), hypertrophic cardiomyopathy (HCM), arrhythmogenic cardiomyopathy (ARR), tetrology of fallot (FALL), inter-atrial communication (CIA), and tricuspid regurgitation (TRI). See the paper for more details. The segmentation labels of LV, RV, and MYO are provided for ED and ES phases for short-axis and long-axis four-chamber views. The images are stored in NIfTI format.

Pathologies and criteria
Abbreviation Name Criteria
DLV Dilated left ventricle LV EDV >214mL(>105mL/m2) for men and >179 mL(>96mL/m2) for women
RV Dilated right ventricle RV EDV >250mL(>121mL/m2) for men and >201 mL(>112mL/m2) for women
HCM Hypertrophic cardiomyopathy LV wall thickness >15mm
ARR Arrhythmogenic cardiomyopathy global RV dilatation and wall motion abnormalities with or without a decreased EF
FALL Tetrology of fallot a nonrestrictive ventricular septal defect, overriding aorta; right ventricle outflow tract obstruction and/or branch pulmonary artery stenosis; and RV hypertrophy
CIA Inter-atrial communication RV volume overload, identification of inferior sinus venous defect in the long-axis 4-chamber view
TRI Tricuspid regurgitation one or more flow jets emanating from the tricuspid valve and projecting into the RV, often holosystolic and readily apparent on the long-axis 4-chamber view

Download Dataset

Note

It is recommended to download the dataset to download the data under ~/.cache/cinema_datasets/mnms2 as the integration tests uses this path. Otherwise define the path using environment variable CINEMA_DATA_DIR.

Download the dataset from the website and unzip the .rar file.

After unzipping the file, the files will have the following structure. Note, 160 training subjects, 40 validation subjects and 160 testing subjects are ordered sequentially.

MnM2.zip
MnM2/
├── dataset_information.csv
├── dataset/
│   ├── 001/
│   │   ├── 001_LA_CINE.nii.gz
│   │   ├── 001_LA_ED_gt.nii.gz
│   │   ├── 001_LA_ED.nii.gz
│   │   ├── 001_LA_ES_gt.nii.gz
│   │   ├── 001_LA_ES.nii.gz
│   │   ├── 001_SA_CINE.nii.gz
│   │   ├── 001_SA_ED_gt.nii.gz
│   │   ├── 001_SA_ED.nii.gz
│   │   ├── 001_SA_ES_gt.nii.gz
│   │   ├── 001_SA_ES.nii.gz
│   ├── 002/
│   ├── ...

Preprocessing

The preprocessing is performed on ED and ES images with the following steps:

  • resampling LAX 4C images to 1.0 x 1.0 mm and SAX images to 1.0 x 1.0 x 10.0 mm.
  • cropping the images to 256 x 256 for LAX and 192 x 192 for SAX, based on LV center at ED frame.
  • normalizing the values to [0, 1].
mnms2_preprocess

The CLI requires to install the cinema package at the root of repository with pip install -e ..