Automated selection of echocardiographic views using a novel artificial intelligence software

Evaluation of a single echocardiographic study involves sorting, selecting, measuring and interpreting a large number of images. As these steps are currently performed manually, this can turn into a repetitive, time-consuming and tedious task. Although the use of artificial intelligence (AI) software could potentially improve this entire workflow, most of the efforts have been focused on measurement automatization. Incorporation of automated selection of echocardiographic views could further optimize this process.

We investigated whether an AI-derived measurements software can be supplemented by automated view selection using an AI model. We hypothesized that automated selection of echocardiographic views yields similar results compared to manually selected views when using the same tool for automated measurements in both groups (Figure 1, top).

We trained an AI model (convolutional neural network) to recognize the main standard 2D echocardiographic views. This model was trained on individual frames from 20,000 DICOM clips to predict a specific cardiac view, based on the view annotation provided by an echocardiography expert (ground-truth). We then selected a test set of 5,000 DICOMs where we applied the AI model to all images in the study. We selected the highest-ranking apical 2 chamber (A2C) and apical 4 chamber (A4C) images, based on the model’s output probability (Figure 1, bottom). For the same 21 studies, an expert selected the best A2C and A4C image for left ventricular volumes and ejection fraction measurements. Then, we ran Philips AutoMeasure both on manually and automatically selected images. We then compared the measurement values for statistical significance using a pair-wise t-test, see Figure 2.

The AI model achieved an accuracy of 96% and AUC of 0.99 on the test set. The results obtained when comparing the automatic and manually selected images in the 21 studies are shown in Figure 2. For all the measured parameters, there was no statistically significant difference in mean values between both selection methods.

Manual and AI-derived selection of standard echocardiographic images do not lead to a statistically significant difference in automated measurements, suggesting that incorporation of automated view selection may contribute to improve daily echocardiographic workflow. Future research could aim in extending these findings to more cardiac views, parameters, and populations.

Comparison of the approaches