---

About

Model Facts

Summary

We provide the user with a deep learning Algorithm for automatic and robust segmentation of the cardiac chambers and myocardium in CCTA images . This method can be used to accurately derive volumes of these cardiac structures in different cardiac phases. The code that we used to train and test the Algorithm is open-source and available on GitHub. Consider uploading your data to this Archive, instead of directly to the Algorithm. The Archive will automatically feed your data to the Algorithm, and it will allow you to send the outputs to this reader study. There you can grade the quality of the segmentation output, correct if needed (to be implemented), and finally, download the results.

Mechanism

We use a 3D convolutional neural network (CNN) architecture for automatic multi-class segmentation of the LV myocardium, LV cavity, RV, LA, and RA across the entire cardiac cycle. The 3D CNN architecture consists of an encoding path with two downsampling layers with strided convolutions, six residual ResNet blocks, and a decoding path composed of two upsampling layers with transposed convolutions. 3D batch normalization is used for stabilized network training. The output layer is a softmax function that predicts per-class probabilities that sum up to 1.

Inputs and Outputs: The Algorithm accepts 3D cardiac CCTA images in .mhd or DICOM format, and gives as output a 3D multi-class mask.

Target population: Any patient undergoing a contrast-enhanced CT scan of the heart, for example in TAVI treatment planning.

Validation and Performance

Shown below are automatic segmentations of the cardiac structures in end-systolic (ES) and end-diastolic (ED) phases of two different patients in the leave-one-patient-out cross validation. a) Very accurately segmented images, b) images with very low contrast enhancement in comparison with the data in the training set and poor automatic segmentation. All images are displayed with the same window-level setting. For a thorough discussion of performance (including metrics), we would like to refer to the journal and conference paper listed at the top of the page.

Uses and Directions

This algorithm was developed for research purposes only. Please be aware that the uploaded images and results will be visible to the owner of the algorithm. Other users won't be able to see your images and results, unless you set them to public. Compatible dataformats include .mhd and .nii.gz -- DICOMs will need to be converted to one of these formats.

Warnings

This algorithm is intended for use with contrast-enhanced cardiac CT only. For whole-heart segmentation of noncontrast CT (NCCT), please see this algorithm.

Common Error Messages

Information on this algorithm has been provided by the Algorithm Editors, following the Model Facts labels guidelines from Sendak, M.P., Gao, M., Brajer, N. et al. Presenting machine learning model information to clinical end users with model facts labels. npj Digit. Med. 3, 41 (2020). 10.1038/s41746-020-0253-3