Cardiac MRI (3-Chamber View)
PhD project by: Rob Hayward
Tim Cootes Chris Miller Tim Morris
The aorta, carrying oxygenated blood from the heart, is susceptible to a number of diseases including bicuspid aortic valve (BAV) disease. Possible complications include aortic dissection and rupture, which are associated with high mortality rates. Various surgical interventions are available, depending on the specific pathology, but all carry significant risk. Current clinical recommendations call for regular monitoring with risk assessment determining intervention.
CT, MRI, and Ultrasound are all useful modalities of aortic imaging; MRIs are commonly used for ongoing monitoring as they allow for accurate measurements without exposure to ionizing radiation. Standardized measurements of aortic diameter, typically made at the sinuses if Valsalva and various points in the aortic arch, are used as a principal metric in determining whether surgical intervention is necessary. Surgery is recommended for BAV patients with a maximal aortic diameter ≥55mm, though a lower threshold of 50mm can be applied in the presence of additional risk factors, such as an increase in diameter >3mm in a year.
There are limitations to the current approach. Aortic measurements are time-consuming and subject to inter-observer variation. There is no consensus on various aspects of the measurements, such as whether the aortic wall should be included or excluded in the aortic diameter measurements or whether measurements done during systole or diastole provide better prognostic information. Furthermore, risk stratification remains poor. The currently used aortic measurements are widely recognised to be inadequate indicators for surgical intervention. Parameters that provide more discriminatory and personalised risk stratification are required.
The goal of this project is to automate the analysis of the aorta from cardiac MRI images. As a first step, this will involve extracting the aortic measurements currently made by human observation. The approaches developed will locate the clinical landmarks and structures and enable determination of novel parameters that describe the shape and appearance in more detail, including changes observed over the cardiac cycle and over time. The goal is to use these parameters along with clinical variables to improve the efficiency and prognostic utility of the scans and aid clinical decisions.
Cardiac MRI (3-Chamber View)
Many approaches have been applied to segment medical images. Current state-of-the-art methods tend to rely on convolutional neural nets, though active shape models can also yield impressive results and tend to outperform CNNs when the number of training examples is limited. Both of these approaches are useful in this project.
Active shape models define a object in terms of labelled points. A statistical shape model is used to describe the relationship between the points, as seen in the training examples, and principal component analysis used to determine the main axes of variation. The statistical shape model then constrains the shape of the active shape model, which iteratively deforms to fit an example of the object in a new image. This is applied by sampling regions of the image around the current estimate and generating a ‘response image’, giving a cost for having the point at each pixel. In a random forest implementation, regressors are trained independently for each model point, and each tree trained on patches sampled at many random displacements. During matching, each patch gives multiple votes for the point position and the votes are accumulated to yield a response image.
This approach has been tested on a training set of 60 3-chamber cardiac MRI images representing 20 patients, with an additional 15 images (representing 5 patients) held for validation. Each image was manually annotated with 17 points. The most significant modes of variation in the shape model are shown below:
 |
 |
 |
|---|---|---|
| Mode 1 | Mode 2 | Mode 3 |
Preliminary results on the validation set show that the model is able to accurately predict the points.
| Ground truth | Prediction | ||
|---|---|---|---|
 |
|||
Additional work and refinement is required with more sophisticated and clinically relevant annotation and with a larger dataset. The accuracy of the model needs to be tested and validated against human annotations of clinical landmarks.
CNNs have been demonstrated to be well-suited for both keypoint regression and segmentation.
As a preliminary test of keypoint regression, the same 60 training images and 15 validation images used with active shape models were used to train and validate a convolutional neural network. A 34-layer ResNet architecture was used with a mean-squared-error loss function.
| Ground truth | Prediction | ||
|---|---|---|---|
 |
|||
Although the model is able to learn some relevant features, the accuracy is far lower than the shape model using the same amount of information. Additional work is required to refine the model and increase the amount of training data.
To investigate CNN-based segmentation, the same images were manually segmented into three categories. These were then used to train a CNN using a U-Net architecture with a cross-entropy loss function. The preliminary results are shown below.
| Ground truth | Prediction | ||
|---|---|---|---|
 |
|||
Future work will concentrate on refining the model and training with additional data.
The utility of scan analysis parameters as prognostic factors (for a combined outcome of increase in aortic size, aortic intervention and death) will be evaluated using survival analysis methodology. Multivariable prognostic models, incorporating clinical variables (such as diagnosis [e.g. syndrome connective tissue disorders and non-syndrome disorders] and other risk factors [e.g. hypertension]) and scan analysis parameters will be developed.
Authors/Task Force members, Erbel, R., Aboyans, V., Boileau, C., Bossone, E., Bartolomeo, R.D., Eggebrecht, H., Evangelista, A., Falk, V., Frank, H. and Gaemperli, O., 2014. 2014 ESC Guidelines on the diagnosis and treatment of aortic diseases: Document covering acute and chronic aortic diseases of the thoracic and abdominal aorta of the adult The Task Force for the Diagnosis and Treatment of Aortic Diseases of the European Society of Cardiology (ESC). European heart journal, 35(41), pp.2873-2926.
Lindner, C., Bromiley, P.A., Ionita, M.C. and Cootes, T.F., 2014. Robust and accurate shape model matching using random forest regression-voting. IEEE transactions on pattern analysis and machine intelligence, 37(9), pp.1862-1874.
He, K., Zhang, X., Ren, S. and Sun, J., 2016. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition (pp. 770-778).
Ronneberger, O., Fischer, P. and Brox, T., 2015, October. U-net: Convolutional networks for biomedical image segmentation. In International Conference on Medical image computing and computer-assisted intervention (pp. 234-241). Springer, Cham.