Cardio Thoracic Ratio for Congenital Heart Disease detection using Deep Learning

Congenital Heart Defect (CHD) is one among the most commonly occurring congenital defects. Prenatal diagnosis of CHD is critical for timely medical intervention and counseling thereby reducing mortality rates. However, chances of misdiagnosis of CHD are high owing small size of the fetal heart and complexity of imaging fetal heart due to the noisy nature of the ultrasound images. The present study focuses on assisting the clinicians in the detection of CHD by advancing techniques for recognition of fetal heart defects using deep learning.This study proposes a computer aided diagnostic tool for the measurement of Cardio Thoracic Ratio (CTR) that is indicative of the occurrence of CHD. It involves developing a deep learning model that is used to identify the cardiac and thoracic regions automatically and measure the CT ratio. This reduces the inter and intra observation errors that are commonly involved in the manual estimation of CT ratio by clinicians. VGG annotation tool is used to annotate the cardiac and thoracic regions with the assistance of clinicians and is used to train the deep learning model based on Mask RCNN. The deep learning model developed is observed to assist in automated identification of the region of interest with an average accuracy of 99.2%. The regions extracted by the deep learning model are also compared with the ground truth of the cardiac and thoracic region using commonly used segmentation evaluation metrics such as Dice coefficient, Jaccard ratio, Structural Similarity Index (SSIM). The mean values of the evaluation metrics indicate that the regions extracted by the model closely matches the ground truth. This is used to compute the CT ratio followed by the classification of CHD or no CHD based on the measurement.