A hybrid neural system for the automatic segmentation of the interventricular septum in echocardiographic images

Abstract

Echocardiographic exams allow the observation and extraction of measures related to cardiac structures. In the longitudinal parasternal view, these measures include the left ventricle end-diastolic and end-systolic diameters, end-diastolic interventricular septum thickness (IVSd), and end-diastolic left ventricle posterior wall thickness (LVPWd). Among these measures, the IVSd is important for diagnosing pathologies like hypertrophic cardiomyopathy, aneurysms, abnormal movement and structural faults. This work presents a hybrid neural network system to segment interventricular septum in echocardiographic images of parasternal longitudinal view. The hybrid system developed here consist of a Self-Organizing Map and a Multilayer Perceptron (MLP) neural network. The approach has two phases: clustering and classification. First, the Self-Organizing Map clusters image patches that are previously labeled as Septum and Non-septum. Later, an MLP is trained with information generated by the map. The MLP is then employed to classify patches of a new image resulting in a mask that indicates the probable septum regions. To validate the results, we did a semi-automatic extraction of septum thickness. The average error between the septum thicknesses obtained by the algorithm and the one manually traced was 0.5477mm +/- 0.5277mm. Future recommendations are presented to improve the hybrid system performance to get more accurate results.