Study and identification of early fetal heart chambers from ultrasonic images using swarm intelligence techniques

Abstract

Congenital Heart Defect (CHD) affects the structure and functions of the early fetal heart, affecting 0.8% of all live births. About one third to one half are severe and lethal unless early intervention is made.The cardiac malformations are detected using ultrasound heart examination during the second trimester of pregnancy. The risk factors are due to maternal risk factors and fetal risk factors. Fetal Echocardiography is a diagnostic tool that detects 90% of the cardiac arrhythmia cases. The Four Chamber View images the four chambers of the fetal heart, namely left ventricle, left atrium, right atrium and right ventricle successfully. The Ultrasound cine loop sequences captured are two dimensional in nature. Though the three dimensional, four dimensional and colour Doppler scans are available, 2-D scan is affordable by the rural people. The early detection of CHD is imperative to avoid and reduce the fetal mortality rate. The four chamber view is currently being assessed by the clinicians, radiologists or obstetricians by the visual inspection, experience and skill as there is no automatic detection method. The limitations of this manual assessment are: it is operator dependent, it is time consuming task and it has poor repeatability. The basic cardiac examination and extended basic cardiac examination may be suggested. The standard examination is the four chamber view and the extended basic examination may involve the evaluation of the two outflow tracts in addition to four chamber view. This research work aims at developing an algorithm to detect the chambers of the fetal heart from the ultrasound videos. 12 videos were used for this study. All these videos were converted into frames at a rate of 24fps. The frames were analyzed and then subjected to Region based Chan-Vese Segmentation. It involved the selection of a single initial contour that evolves based on the level set evolution equation to detect the Region of Interest. The performance of the region based CV model segmentation was evaluated and recorded. The average number of iterations was 200 optimum. Though 200 was the optimum value, the segmentation process was incomplete and unsuccessful. Also, due to improper initial contour selected, the solution got stuck in a local minimum rather than the global minimum. So, a novel approach that combines Swarm Intelligence based Cat Swarm Optimization (CSO) technique and Nature inspired Flower Pollination Algorithm (FPA) was used to find the optimal contour that evolved as the best. This method was called GPCATS based Region based Chan-Vese model. The Swarm Intelligence based Cat Swarm Optimization was combined with Flower Algorithm. The CSO operated in seeking mode when cats were resting and in tracing mode when cats were chasing their prey. CSO simulated the behavior of cats when they were at rest observing the surrounding and when they were in movement chasing prey after resting. In FPA, the pollination characteristics of flower were simulated. It involved global pollination and local pollination. In this research work, the global pollination step of FPA was used to improve the distance averaging of CSO. The GPCATS optimiser used many initial contours unlike traditional Chan-Vese where only one initial contour was used. The contours initialized would be made to evaluate the fitness function newline