Machine learning much faster than humans at reading heart images
Automated AI analysis of cardiac magnetic resonance images can be just as accurate as trained clinicians—but 186 times faster.
Automated AI analysis of cardiac magnetic resonance images can be just as accurate as trained clinicians—but 186 times faster.
Researchers say the technique can improve evaluations and reduce inconsistencies in interpretation.
Cardiovascular MRI is the reference standard to assess the structure and function of a heart’s left ventricle, a key imaging biomarker used for clinical decision making and as a clinical trial outcome measure.
However, clinical analysis of these MRIs is significantly variable. The analysis is often not standardized, and there are avoidable inconsistencies when using human interpreters, such as noise or bias.

The researchers, from several institutions in the UK, theorized that automated techniques using machine learning may offer improved readings of the heart, but that the results should proven before there’s widespread adoption.
They trained a fully automated convolutional neural network to read left ventricular chamber volume, mass and ejection fraction from MRIs. They then pitted the model against an expert clinician and two trained junior clinicians in reading 110 scan-rescans of patients at five institutions. Some of the patients had various cardiovascular diseases; others were healthy subjects.
There was no significant difference in the accuracy of the readings between the expert, the trainees and the automated analysis for all of the heart measures.
However, the clinicians took an average of 13 minutes for analysis per scan, compared to about four seconds for the automated model. It also took only nine hours to train the model; it took a month to train the junior trainees.
“The adoption of [machine learning] can offer comparable precision with clinicians, with the time saving and global standardization that would ensue….In the UK, an estimated 2,275 scans per million adults are needed annually, performed in 61 centers. Automating this one aspect of [cardio magnetic resonance] analysis alone would, therefore, potentially translate into a saving of 54 clinician-days per center. Accurate automated segmentation is a bridge to reliable extraction of more information from the same imaging beyond established imaging biomarkers. In combination with time saving, this maximizes use of acquired data in a value-based manner,” the study authors noted.
The study was published in the American Heart Association’s journal Circulation: Cardiovascular Imaging.
“Given that the greatest sources of measurement error were human factors … we believe that, with improvement, it is only a matter of time before automated approaches are super-human, with cascading consequences in clinical … and research … domains of increased precision,” the researchers stated.
Researchers say the technique can improve evaluations and reduce inconsistencies in interpretation.
Cardiovascular MRI is the reference standard to assess the structure and function of a heart’s left ventricle, a key imaging biomarker used for clinical decision making and as a clinical trial outcome measure.
However, clinical analysis of these MRIs is significantly variable. The analysis is often not standardized, and there are avoidable inconsistencies when using human interpreters, such as noise or bias.

The researchers, from several institutions in the UK, theorized that automated techniques using machine learning may offer improved readings of the heart, but that the results should proven before there’s widespread adoption.
They trained a fully automated convolutional neural network to read left ventricular chamber volume, mass and ejection fraction from MRIs. They then pitted the model against an expert clinician and two trained junior clinicians in reading 110 scan-rescans of patients at five institutions. Some of the patients had various cardiovascular diseases; others were healthy subjects.
There was no significant difference in the accuracy of the readings between the expert, the trainees and the automated analysis for all of the heart measures.
However, the clinicians took an average of 13 minutes for analysis per scan, compared to about four seconds for the automated model. It also took only nine hours to train the model; it took a month to train the junior trainees.
“The adoption of [machine learning] can offer comparable precision with clinicians, with the time saving and global standardization that would ensue….In the UK, an estimated 2,275 scans per million adults are needed annually, performed in 61 centers. Automating this one aspect of [cardio magnetic resonance] analysis alone would, therefore, potentially translate into a saving of 54 clinician-days per center. Accurate automated segmentation is a bridge to reliable extraction of more information from the same imaging beyond established imaging biomarkers. In combination with time saving, this maximizes use of acquired data in a value-based manner,” the study authors noted.
The study was published in the American Heart Association’s journal Circulation: Cardiovascular Imaging.
“Given that the greatest sources of measurement error were human factors … we believe that, with improvement, it is only a matter of time before automated approaches are super-human, with cascading consequences in clinical … and research … domains of increased precision,” the researchers stated.
More for you
Loading data for hdm_tax_topic #better-outcomes...