Visualisation of a 3D bi-ventricular model obtained through segmenting the volumetric image.
Imaging Latest News News Other Categories

Artificial Intelligence, precision phenotyping and risk prediction in cardiomyopathy

CardioNewsUK 2 July 2021

By Ahmed El-Medany

Presented by Prof O’Regan, MRC Investigator and Consultant Radiologist, Imperial College London

The Digital Innovation in Cardiology session at BCS 2021 was a fascinating and inspiring insight into the technology available in various areas of Cardiology and covered exciting digital prospects for the future.

Prof O’Regan discussed Artificial Intelligence’s (AI) role in cardiovascular imaging; particularly when integrating imaging and non-imaging data, understanding new mechanisms of heart disease, and early diagnosis and risk prediction.

The speaker demonstrated some excellent examples of 3D segmentation cardiovascular MRI[1] (Figure 1).

‘Typically we rely on crude, quantitative measures of volumes and function’ Prog O’Regan stated. ‘A lot of this work is trying to make best use of the images of the data we already have’.

‘We can very readily and accurately scale imaging analysis to big cohorts and across different types of diseases…

We can also develop these very rich three-dimensional motion models of the heart, which allow us to capture the nature complexity of cardiovascular disease, that we may not be able to do with conventional measures’.

Figure 1: Taken from Prof O’Regan’s presentation at BCS 2021. Visualisation of a 3D bi-ventricular model obtained through segmenting the volumetric image.

The presenter then showed an example of AI automated diagnosis based on 3D shape and motion, using a hypertrophic cardiomyopathy (HCM) model. ‘As clinicians looking at the image, we look at the whole image… we don’t just rely on the simple linear metric’.

An ‘auto-encoder’, a form of data compression, can be used to discriminate between different 3D heart models to reconstruct images specific to each cardiomyopathy (Figures 2 and 3). ‘[Automated diagnosis] captures, implicitly, differences in strain and regional differences in wall thickness’ the Professor stated. Moreover, this information can lead to us understating how the heart changes in shape and geometry, and the spectrum between health and disease [2].

Figure 2: Taken from Prof O’Regan’s presentation at BCS 2021. Originally adapted from Biffi et al (2020). Use of a discriminative network to predict cardiomyopathy type.
Figure 3: Taken from Prof O’Regan’s presentation at BCS 2021. Originally adapted from Biffi et al (2020). Example automated discrimination based on complex 3D shape and motion traits Healthy volunteers (HVol) in green and hypertrophic cardiomyopathy (HCM) in red.

With regards to predicting outcomes in disease Prof O’Regan described a deep learning algorithm to automatically segment pulmonary hypertension MRI images, and predict survival based on motion of the right ventricle (RV) and interventricular septum (IVS).

Pixel precise analysis via AI and analysis of 300 points on the surface of the RV can create a digital model of how the RV moves in systole and diastole. Subsequently, aggregation of data from several hundred participants can be used to describe the variation in structure and function of the RV in a large cohort of patients.

Again, an auto-encoder can be used to find simple patterns in this complex data, useful for predicting survival in pulmonary hypertension (Figure 4) – a so-called ‘4DSurvival Network’ [3] .

Prof O’Regan stated: ‘Pulmonary hypertension outcomes at 1 year determine guideline driven management.. this can be really important for ensuring patients receive the right treatment’.

Figure 4: Taken from Bello et al, 2019. Deep learning cardiac motion analysis for survival prediction.

Next, the speaker discussed how imaging and genetics could be integrated using AI. 3D phenotype mapping (Figure 5) was performed in over 2,000 healthy individuals. Aggregated data was collected to look for truncating variants in Titin, a protein strongly implicated in dilated cardiomyopathy (DCM), which is also seen in around 1% of the healthy population[4,5].

Figure 5: Taken from Prof O’Regan’s presentation at BCS 2021. Originally adapted from Biffi et al (2019) and Schafer et al (2017). Computational modelling of cardiac geometry in healthy humans using 3D CMR. Positive standardized β coefficients indicate where TTNtv status is associated with enlargement of the left ventricle cavity at end diastole (left) and end systole (right).

3D phenotype mapping (Figure 5) has shown that truncating variants have expressivity in healthy heart, although may not always lead to overt DCM, unless exposed to alcohol, chemotherapy, or pregnancy.

‘This can be a powerful way to integrate genetic data with imaging’ the presenter stated.

References

  1. DP, Rueckert D. Automatic 3D bi-ventricular segmentation of cardiac images by a shape-refined multi-task deep learning approach. IEEE transactions on medical imaging. 2019 Jan 23;38(9):2151-64.
  2. Biffi C, Cerrolaza JJ, Tarroni G, Bai W, De Marvao A, Oktay O, Ledig C, Le Folgoc L, Kamnitsas K, Doumou G, Duan J. Explainable anatomical shape analysis through deep hierarchical generative models. IEEE transactions on medical imaging. 2020 Jan 6;39(6):2088-99.
  3. Bello GA, Dawes TJ, Duan J, Biffi C, De Marvao A, Howard LS, Gibbs JS, Wilkins MR, Cook SA, Rueckert D, O’regan DP. Deep-learning cardiac motion analysis for human survival prediction. Nature machine intelligence. 2019 Feb;1(2):95-104.
  4. Biffi C, de Marvao A, Attard MI, Dawes TJ, Whiffin N, Bai W, Shi W, Francis C, Meyer H, Buchan R, Cook SA. Three-dimensional cardiovascular imaging-genetics: a mass univariate framework. Bioinformatics. 2018 Jan 1;34(1):97-103.
  5. Schafer S, De Marvao A, Adami E, Fiedler LR, Ng B, Khin E, Rackham OJ, Van Heesch S, Pua CJ, Kui M, Walsh R. Titin-truncating variants affect heart function in disease cohorts and the general population. Nature genetics. 2017 Jan;49(1):46-53.

Links
Full session available On Demand: https://bcs2021.co.uk

Automatic 3D Bi-Ventricular Segmentation of Cardiac Images by a Shape-Refined Multi- Task Deep Learning Approach | IEEE Journals & Magazine | IEEE Xplore

Explainable Anatomical Shape Analysis Through Deep Hierarchical Generative Models | IEEE Journals & Magazine | IEEE Xplore

Deep-learning cardiac motion analysis for human survival prediction | Nature Machine Intelligence

Three-dimensional cardiovascular imaging-genetics: a mass univariate framework | Bioinformatics | Oxford Academic (oup.com)

Titin-truncating variants affect heart function in disease cohorts and the general population | Nature Genetics