Heart ultrasound examination is heavily loaded with manual work. Exam takes from 30 to 90 minutes to complete and a physician would normally do around 250 clicks during the exam. Our solution recognises different heart image views and performs required measurements without human input. Speeds up the examination 6 times and decreases manual work for physicians.

Value Proposition

Pain points in heart ultrasound examination: 85% of total examination time is spent on doing manual measurements, on average there will be 250 keyboard clicks made during the exam, the examination time averages from 30 to 90 minutes, there is a great inter-operator variability (when two different physicians assess a same patient, they get different results), and patients wait on average 5 weeks to get admitted to a hospital. This solution automates measurements otherwise done manually during a regular heart ultrasound examination. This allows clinicians to finish the examination in 5 minutes instead of 30. This is particularly important in the times of COVID-19 when we must limit physicians' exposure to patients. Moreover, approximately 30% of patients positive for COVID-19 will have an underlying cardiovascular pathology and the first method of choice for differential diagnosis and patient management is heart ultrasound examination. Therefore, this solution is as important as ever.

Narrative

Currently, 2D echocardiography is performed in a hospital by either a cardiologist or a technician. Some countries e.g. the Netherlands, the USA have the technicians who perform the examination whereas in other countries e.g. Germany, Lithuania echocardiography is performed by cardiologists. Anyways, 2D echocardiography takes from 30 to 90 minutes to complete. According to the guidelines of European Society of Cardiology there are 50 different measurements to be done in 9 different image views. When performing each measurement an operator (a physician or a technician) must freeze echocardiography images at certain points called end-systole and end-diastole and make each of those measurements. This sounds (and actually is!) a very tiring job because everything is done manually. In case of the Netherlands and the USA, images and measurements from ultrasound machine performed by a cardiologist are sent to a workstation of a cardiologist for the so-called post-procesing of images. A cardiologist then validates everything what was previously annotated (or not) by a technician and signs the report. In other countries, cardiologists jump to their workstation if they missed or could not properly perform measurements during the examination and only then annotates images and signs under a report. Our solution allows to skip manual measurements because they will be done by our neural network. This will allow both technicians and cardiologists speed up the examination time 6-fold. The only job for cardiologists will be validation of what was already measured by our tool.

Intended Use

Intended use: analysis of DICOM format 2D transthoracic echocardiograms. Patients include those above 18 years old of either gender. The software cannot be used on congenital heart malformations. Indications of use include but are not limited to various types of heart failure, myocarditis, valve pathology, hydropericardium etc.

Limitations

The model was trained on primarily caucasian patients. The training dataset comes from the Netherlands, Lithuania and Latvia, therefore is considered to be homogenous. The model was not trained on children and on patients having congenital heart malformations.

Model performance metrics

We measure standard error of measurement (SEM) for every single measurement in ever single heart-image view. This represents accuracy against human-operators. With our currently rather small dataset 87% of all the parasternal long-axis measurements fall under the SEM 95% confidence interval.

LigenceLogo1

UAB LIGENCE

UAB LIGENCE was founded by medical specialists back in 2019. Team members have also experience in deep learning projects thus they decided to tackle various health care system problems using deep learning. So far, UAB LIGENCE has secured a pre-seed funding round of 360k and has been granted two EU-backed projects of which one is dedicated to help fight COVID-19. Our ultimate mission is to relieve doctors from repetitive tasks and empower them with technology.