Computer Vision News - February 2022

33 Automated Echocardiographic Detection of ... To pursue this, different ML classifiers are employed: support-vector machine, random forest and logistic regression models are all trained and compiled in a soft voting strategy. Stratified and repeated 2-fold nested cross validation is performed for hyperparameters optimization and feature selection. Model predictions are finally compared to the ground truth outcomes at a range of class prediction probability thresholds. Moreover, the models are also independently tested using a retrospective study of patients who had undergone SE and had follow-up to at least 6 months through clinical record review. All images are processed using the same automated AI pipeline and then the model is used to identify patients with severe CAD. The second contribution of the paper is instead a direct clinical translation: This question is answered by the authors through a multiple reader multiple case (MRMC) randomized crossover design study. This uses medical records of 148 SE studies which were selected to be suitable from the test set of the previous study. Four experienced physicians/ echocardiographers who were independent of any other part of the investigation were selected to classify CAD. For each reader, 50% of the studies were randomly provided with a report stating if the AI-based classifier deemed patients as having severe coronary artery disease. After a 1-month wash out period, the reader was presented with the same images but on this occasion AI-based classification was provided for the alternative 50% of patients. Results Let’s now look at how the authors managed to answer these issues. First, it’s worth showing part of the pre-processing pipelinewhich does the auto-contouring (red) accuracy against manual contours (teal) from BSE accredited echocardiographers in contrast enhanced and non-contrast images in the apical 2-chamber, apical 4-chamber and parasternal short axis at the level of the mitral valve views. The accuracy is indicated at the bottom of each image and provides confidence that the future steps of feature extraction and disease classification can be properly performed. Now, getting into the novel findings from the paper, the feature selection process discovered 31 features, whose individual evaluation demonstrated efficacy in identifying patients with severe disease. On the left of the image below, you can observe two plots which show the capability of individual model features to differentiate outcome where the vertical and horizontal lines indicate example cut- off values for disease classification, and the red and green dots identify CAD negative and positive patients. Do clinicians benefit if provided an automatic classification?