Overview

Schema Version

https://atlas.rsna.org/schemas/2025-11/model.json

Name

Left-Ventricle Quantification Using Residual U-Net

Link

https://github.com/ericspod/STACOM18

Indexing

Keywords: Cardiac MR, Cardiac Quantification, Convolutional Neural Networks

Content: CA, MR, BQ, CH, IN

Author(s)

Eric Kerfoot

James Clough

Ilkay Oksuz

Jack Lee

Andrew P. King

Julia A. Schnabel

Organization(s)

King's College London

NVIDIA Corporation

Funding

Supported by an EPSRC programme Grant (EP/P001009/1) and the Wellcome EPSRC Centre for Medical Engineering at the School of Biomedical Engineering and Imaging Sciences, Kings College London (WT 203148/Z/16/Z). GPUs generously donated by the NVIDIA Corporation.

Model

Architecture

U-net convolutional neural network architecture built from residual units. Uses parametric rectifying linear units and instance normalization. Employs convolutions with a stride of 2 for downsampling and transpose convolutions with a stride of 2 for upsampling.

Availability

Code available to download at https://github.com/ericspod/STACOM18.

Clinical benefit

Provides diagnostic values for assessing cardiac health and identifying pathologies, offering time and cost savings for clinical diagnostic use.

Clinical workflow phase

Diagnosis

Degree of automation

Fully automated

Indications for use

Quantification of left ventricle metrics (cavity and myocardium volumes, cavity dimensions, regional wall thicknesses, and assignment to diastolic or systolic portion of the cardiac cycle) for assessing cardiac health and diagnosing pathologies such as hypertrophic cardiomyopathy.

Input

Full cycle cardiac MR images (cMRI) captured at three positions on the LV, with input image matrix dimensions of 2900 × 80 × 80 pixels.

Instructions

Trained using PyTorch 0.4.0 on nVidia TITAN Xp and P6000 GPUs. Training takes approximately 3 hours for 10000 steps with a batch size of 1200. Segmentation analysis routines are implemented in Python 3.6 using Numpy, SciPy, and Numba libraries.

Limitations

Meaningful loss of accuracy on the test dataset, with some images not segmented correctly, requiring compensation for incomplete annular segmentation. Poor results can occur with images dissimilar from training data, or when image noise makes the myocardium indistinct.

Output

Description: Myocardial segmentations of the left ventricle, from which cavity and myocardium volumes, cavity dimensions (IS-AL, I-A, IL-AS), six regional wall thickness values (IS, I, IL, AL, A, AS), and assignment to diastolic or systolic portion of the cardiac cycle are derived.

Reproducibility

The network is generally stable and insensitive to varying training input datasets, with IoU values within folds being roughly equivalent. Data augmentation is crucial for generalizability and preventing overfitting.

Use

Intended: Left Ventricle Quantification, Cardiac Health Assessment, Pathology Identification, Image Segmentation

User

Intended: Diagnostic Radiologist, Cardiologist, Researcher, Physician