Gradient Boosting Decision Trees for Echocardiogram Images

High-throughput medical imaging procedures, such as echocardiogram, motivate the development of new computational tools to support specialists in decision making. This paper introduces a method to detect the posterior wall of the left ventricle from echocardiogram images, considering the PLAX view. Using a Light Gradient Boosting Machine, we evaluated a classification problem and compared results with Classification and Regression Trees using k-fold cross-validation regarding five quality measures. Experimental analysis shows that the proposed approach achieves relevant median scores of 61% for Dice, 67% for Sensitivity and 98% for Specificity on a dataset of only 69 images from three different sources. With some post-processing, the median Dice score increased to 78%. Although further improvements must be considered before deploying such a tool for clinical utilization, the results presented in this paper show a promising path towards the automation of echocardiogram analysis and segmentation of the posterior wall of the left ventricle.