Prospective Validation of an Artificial Intelligence Algorithm Identifying Echocardiographic Reference Views. Ultrasound - Cardiac Acquisition Guide
Prospective Validation of an Artificial Intelligence Algorithm Identifying Echocardiographic Reference Views. Ultrasound - Cardiac Acquisition Guide
Echocardiography is the examination of choice for the study of cardiac pathologies. Beyond its use by cardiologists, the interest of echocardiography for other medical specialties has already been demonstrated, in particular in intensive care in the case of haemodynamic failure, or in intra and extra hospital emergency medicine for the initial assessment of chest pain or dyspnoea.
Echocardiography also plays a major role in screening for heart disease, particularly valvular heart disease. In countries with very limited access to echocardiography, there is a major under-diagnosis of heart valve disease, including rheumatic fever, which affects 30 million people and causes 305,000 deaths worldwide. As this is a global public health problem, recommendations were drafted in 2012 to organise and facilitate echocardiographic screening of populations at risk.
The expansion of the use of echocardiography has been catalysed by the miniaturisation of ultrasound systems and the reduction in their price. Recently, probes directly connected to a tablet or phone have been developed at a limited cost.
It is therefore possible to consider these ultrasound scanners as the new stethoscope that could be used by any health professional.
In order to be effective, the last limit to this democratisation is the training, and in particular that of non-specialists (i.e. non-cardiologists).
Echocardiography remains an examination that requires anatomical knowledge and practice. Performing an echocardiogram involves visualising the heart from different points on the chest. The three main points are in the left paraspinal area, at the apex of the heart and under the sternum. From these areas, the operator must obtain several reference views which are strictly defined in order to be able to correctly observe the different cardiac structures and make comparable measurements from one examination and clinician to another.
It is therefore necessary first of all to learn how to handle the probe and to be able to obtain the reference views. The morphology of the patient, the shape of the thorax, the exact position of the heart, the movements of the heart according to the position of the patient and his breathing are all elements to be taken into account and make each examination different from the previous one.
In response to this problem, several teams have taken advantage of advances in deep learning, particularly in the field of computer vision, to help non-specialists obtain these reference views. Using convolutional neural networks, several teams have developed algorithms for identifying and distinguishing these views. The objective of this work is to provide assistance to the operator by identifying in real time the ultrasound image obtained as a reference view.
Since January 2019, the echocardiography laboratory headed by Prof. Stéphane Lafitte and DESKi, a Bordeaux-based start-up specialising in deep learning and medical imaging, have been working on this type of solution. In particular, they were able to develop an algorithm based on retrospective data that classifies images according to 7 reference views (parasternal long axis, parasternal short axis, apical 4-3-2 cavities, sub costal 4 cavities and the inferior vena cava) and a class representing ultrasound images that do not correspond to any of these views.
The particularities of this work lie mainly in the fact that the algorithm only detects views with sufficient quality (echogenicity and visible anatomical part in the image) for a reliable analysis and that the architecture of the neural network is compatible with a real time use on a smart phone.
Independently of the teams, all these algorithms were built and validated from retrospective data, i.e. loops of ultrasound images recorded by cardiologists during standard echocardiography examinations. In these recordings, the cardiologists keep only the images corresponding to the reference views. The cardiologist then attaches each image loop or acquisition to a reference view. From these recordings labelled by the cardiologists and by a learning method, the algorithms learn to detect and distinguish these reference views. The algorithms are then validated on a sample that has not been used for training, by comparing their results with the labelling performed by the cardiologists.
The limitation of this validation is that it takes little account of the behaviour of the algorithm when confronted with images that are not reference views.
Indeed, before recording these reference views, the operator searches for the position of the probe that offers the best view by moving it over the patient's torso. This whole scanning phase is performed during the standard examination without being recorded. None of these solutions has therefore been validated prospectively on acquisitions including the scanning phase of the reference views.
Through a prospective monocentric study, the objective of the research is to compare the labelling carried out by the algorithm and that carried out by cardiologists from acquisitions including the recording of the reference view search phase and obtained as part of routine care in the echocardiography laboratory.
Inclusion Criteria:
Exclusion Criteria: