Automated Recognition of Congenital Heart Defects (CHD) from Ultrasonic Fetal Heart Biometry

Abstract

The growth and development of the fetus is monitored widely using Ultrasound imaging modality and has become synonymous with prenatal scanning. Analysing the images helps in the detection of congenital defects at an early stage which help to plan the delivery appropriately or for taking some corrective measures. The inherent noise of ultrasound imaging makes analysis of the images largely dependent on the expertise of the clinician and is prone to errors. Computer Aided Diagnostic (CAD) tools used in medical diagnosis assist the clinicians in analysing the images and reducing the chances of missing the defect. AI tools, machine learning and deep learning algorithms have shown improved classification performance of medical images in recent times have resulting in widespread implementation of these technologies in CAD tools. The proposed work discusses the use of machine learning for detection of Congenital heart disease (CHD). Frames resulting from the B mode Ultrasound cineloop sequence have been used for the study. A total of 870 frames with 550 normal frames and 320 abnormal frames are used for the study. Law’s texture feature have been extracted from the pre processed frames which in turn is used for classification. The performance of each of these features is compared using different kernels of SVM classifier. Tuning parameters regularization and gamma have been used to optimise the classification performance using grid search algorithm. It is observed that Laws texture feature results in better classification performance with an average accuracy and precision of 95% and specificity, sensitivity of 98%. Precision Recall curves and AUC-ROC plots have been used for comparing the performance of different kernels.