Artificial Intelligence for Reconstruction of Echocardiography Studies

Breast cancer is the second most prevalent cancer worldwide, causing 11.6% of global cancer diagnoses.1 It is a common and significant cause of morbidity and mortality. Despite improving survival, 11,500 women and 85 men die from breast cancer in the United Kingdom annually.2 Chemotherapy with trastuzumab is an important part of treatment but carries a potential risk of chemotherapy-induced cardiotoxicity, observed as clinical or subclinical left ventricular dysfunction.3 The implications of such cardiotoxicity are important, to both the chemotherapy regime as well as to cardiac function. Thus, it is a potential disruptor to the chemotherapy itself, and while it may in some cases produce subclinical cardiac dysfunction, in others it may result in overt heart failure with its associated signs and symptoms.

The early and accurate diagnosis of cardiotoxicity is therefore vital for preventing significant reduction in left ventricular (LV) function, and to afford institution of cardioprotective medications that may both improve cardiac function and even allow a return to chemotherapy with continued heart failure medications on board. Conversely, late detection may result in more serious heart failure, and a lower ability to continue trastuzumab chemotherapy.

In line with current international guidance, serial transthoracic echocardiography (TTE) is the first line investigation of choice for the surveillance of cardiac function.4,5 This requires standard 2-D echocardiography with as high image quality as possible, to accurately quantify left ventricular size and function. In the United Kingdom, minimum datasets for standard echocardiography are outlined by the British Society for Echocardiography (BSE) guidelines, these include standard views and measurements to ideally be obtained or calculated during a scan. For patients having undergone potentially cardiotoxic chemotherapy the BSE and British Cardio-Oncology Society (BCOS) advise measures as outlined below in addition to the minimum echocardiography data set: 5

1. In all cases, volumetric analysis of left ventricular cavity size and function
2. Simpson's biplane method. This is a conventional measure that using 2D imaging, from two orthogonal planes, to estimate left ventricular volumes.
3. Ideally, and subject to image quality, volumetric quantification of cardiac function should be made using 3D LV size and ejection fraction analysis from a 3D dataset acquired using a 3D transducer. This overcomes the mathematical assumptions central to (a) above, although, temporal resolution is sacrificed at the expense of 3D image acquisition.
4. 2-D global longitudinal strain (GLS) analysis. This is highly dependent on image quality.

Volumetric assessment by 3D echocardiography provides the benefit of overcoming assumptions and possible inaccuracies resulting from Simpson's biplane method. A 3D probe and appropriate equipment and software is required, and the use of 3D imaging reduces the temporal resolution. Strain analysis confers a significant benefit of early indication of possible cardiotoxicity however it is highly dependent on image quality, requiring no more than 2 segments with poor endocardial definition to be seen across the standard apical chamber views (A4C/A3C/A2C).5 The limitations of these two measures result in a significant proportion of patients requiring further surveillance imaging with contrast echocardiography or CMR. 6 This comes at a time and cost burden to both the patient and service provider . This is variable from unit to unit depending on individual subjective threshold on what degree of image quality is considered adequate, and in addition, available of more advanced imaging modalities.

The V-echo VMS+ system uses AI with a CMR knowledge-base to reconstruct a 3D model of the heart and quantify 3D cardiac chambers volumes and ejection fraction, a marker of systolic function, from images acquired by conventional echocardiography system. 3D reconstruction is improved by applying geometric tags using magnetic patches to the right shoulder and to the ultrasound transducer, introducing spatial coordinates to integrate with 2D ultrasound images. The Ventripoint VMS+ system (Toronto, Canada) has previously been studied predominantly in the paediatric population in validity studies comparing performance to alternative edge detection tools in conventional echo, contrast echo and CMR. Performance outputs from these studies have included agreement between the two methods and processing time taken. Studies thus far have not included the adult oncology population that this project will involve.7,8,9

The L-echo Ligence Heart system (Kaunas, Lithuania) utilises an ensemble of convolutional neural networks to both identify views and to infer volume and strain data. The L-echo system has previously been studied in the adult population in predominantly retrospective projects aiming to validate accuracy of view identification, echocardiographic volume and strain data. The L-echo system requires no additional hardware and can be utilised as a post hoc software tool. 10

At the patient level, AI echocardiography analysis using the V-echo and L-echo systems have the potential to improve accuracy, reduce time for analysis, reduce the need for additional diagnostic imaging tests where results are indeterminate or spurious, reduce intra- and inter-operator variability both at baseline and serially over time. At the clinical service level, analysis using AI based systems has the potential to save analysis and reporting time, reduce appointment lengths, reduce the need for additional imaging appointments (e.g. contrast-enhanced echocardiography, and/or cardiac MRI), and increase departmental echocardiography capacity.

This study aims to systematically evaluate the feasibility, workflow impact, accuracy and cost-effectiveness of incorporating AI-powered echo analysis and reporting into the diagnostic workflow of left ventricular systolic dysfunction detection in cardio-oncology patients.

In line with current international guidance, serial transthoracic echocardiography (TTE) is the cornerstone in surveillance of cardiac function.4,5 Echocardiography in its standard form incorporates 2D imaging, and 3D acquisition where possible. This is logical, since the heart is a 3D structure, and 3D image acquisition will provide the most accurate representation of its geometric structure.

However, 3D echocardiography assessment is not feasible in all patients, and is achievable in as low as 66% to at best 91% of patients.11,12 This is a major limitation to its application in patients for the surveillance for cardiotoxicity. Similarly, strain analysis requires high quality 2D imaging. A more advanced consideration, that is strain analysis of 3D image acquisitions, is not in the guidelines.13 It is even more challenging and limited by the low temporal resolution and lack of clinical outcome data to inform its utility.14 This remains an area for further research into clinical utility.

There is therefore a gap between guideline ideals and real-world image quality requirements to perform high-quality, 2-D quantitative derivation of volumetric data and strain metrics. There is an even larger gap between real-world clinical 3D image acquisition and its volumetric analysis and guideline ideals.

Clinically this means that many laboratories will acquire images that allow only simple 1-dimensional chamber size assessment, or poor-quality attempts at a biplane analysis that do not meet the image quality criteria set out in international guidance. This presents the potential to cause erroneous metrics, and the possibility to miss cardiotoxicity through overcalling cardiac function, or diagnosing cardiotoxicity when it is not, in fact, present.

Where resources permit, poor 2D image quality can be remedied through the use of ultrasound contrast agents, or, via CMR. Contrast agents, while generally well tolerated, are contraindicated in pregnancy and pulmonary hypertension, and can result in potentially fatal allergy, albeit rarely.15

Whilst echocardiography is widely used as the first imaging modality in cardiac diagnostic pathways, and as noted, whilst suboptimal image quality can be improved with ultrasound contrast agents, cardiovascular magnetic resonance (CMR) imaging is regarded as the gold standard non-invasive method for measuring cardiac chambers volumes and ejection fractions. However, it is time-intense, not suitable for claustrophobic patients, not suitable for patients above a certain size, cannot be used in patients with ferromagnetic metal implants or foreign bodies, and access to it is markedly more limited than is access to echocardiography. This makes it a rather impractical solution for clinical surveillance of cardio-toxicity.

An interesting set of clinical and scientific questions is therefore raised: in the quest for a superior representation of cardiac function using 3D volumetric quantification, could an extrapolated 3D representation of the heart be derived from conventional 2D images through the use of spatial data superimposed onto 2D images? This bears relevance to left ventricle, right ventricle and atrial assessments. Furthermore, is it feasible to compute 3D strain analysis of such a 3D structure, whether from the left or right ventricles or of the atria, and how do these indices relate to conventional 2D longitudinal strain measurements that are contained in guidance documents? Finally, how does such volumetric image acquisition and analysis alter workflow, the need for additional contrast-enhanced echocardiography or CMR, and clinical service cost-effectiveness?

This study aims to systematically evaluate the feasibility, workflow impact, accuracy and cost-effectiveness of incorporating AI-powered echo analysis and reporting into the diagnostic workflow of left ventricular systolic dysfunction detection in cardio-oncology patients.