TY - JOUR
T1 - Automated image segmentation for cardiac septal defects based on contour region with convolutional neural networks
T2 - A preliminary study
AU - Nova, Ria
AU - Nurmaini, Siti
AU - Partan, Radiyati Umi
AU - Putra, Sukman Tulus
N1 - Funding Information:
We thank all our colleagues and all cardiac septal defect patients who contributed greatly to the current study. We greatly appreciate the support of all students in the Intelligent System Research Group (ISysRG), Faculty of Computer Science, Universitas Sriwijaya.
Publisher Copyright:
© 2021 The Authors
PY - 2021/1
Y1 - 2021/1
N2 - Echocardiogram examination is important for diagnosing cardiac septal defects. With the development of AI-based technology, an echocardiogram examination previously performed manually by cardiologists can be done automatically. Automatic segmentation of cardiac septal defects can help a physician to make an initial diagnosis before referring a pediatric cardiologist for further treatment. In previous studies, automatic object segmentation using convolutional neural networks (CNNs) was one of the DL applications that have been developed for cardiac abnormalities. In this study, we propose a CNN-based U-Net architecture to automatically segment the cardiac chamber to detect abnormalities (holes) in the heart septum. In this study, echocardiogram examinations were performed on atrial septal defects (ASDs), ventricular septal defects (VSDs), atrioventricular septal defects (AVSDs), and normal hearts with patients undergoing echocardiogram examination at Moh Hoesin Hospital in Palembang. The results show that even for the relatively small number of datasets, the proposed technique can produce superior performance in the detection of the cardiac septal defects. Using the proposed segmentation model for four classes produces a pixel accuracy of 99.15%, mean intersection over union (IoU) of 94.69%, mean accuracy of 97.73%, sensitivity of 96.02%, and F1 score of 94.88%, respectively. The plots of the loss and accuracy curve show that all the errors were small, with accuracy rates reaching 99.05%, 98.62%, 99.39%, and 98.97% for ASD, VSD, AVSD, and normal heart, respectively. The comparison accuracy of contour prediction for U-Net was 99.01%, while V-Net was 93.70%. This shows that the U-Net has better accuracy than the V-Net model architecture. It can be proven that the architecture of CNNs has been successful in segmenting the cardiac chamber to detect defects in the heart septum and support the work of cardiologists.
AB - Echocardiogram examination is important for diagnosing cardiac septal defects. With the development of AI-based technology, an echocardiogram examination previously performed manually by cardiologists can be done automatically. Automatic segmentation of cardiac septal defects can help a physician to make an initial diagnosis before referring a pediatric cardiologist for further treatment. In previous studies, automatic object segmentation using convolutional neural networks (CNNs) was one of the DL applications that have been developed for cardiac abnormalities. In this study, we propose a CNN-based U-Net architecture to automatically segment the cardiac chamber to detect abnormalities (holes) in the heart septum. In this study, echocardiogram examinations were performed on atrial septal defects (ASDs), ventricular septal defects (VSDs), atrioventricular septal defects (AVSDs), and normal hearts with patients undergoing echocardiogram examination at Moh Hoesin Hospital in Palembang. The results show that even for the relatively small number of datasets, the proposed technique can produce superior performance in the detection of the cardiac septal defects. Using the proposed segmentation model for four classes produces a pixel accuracy of 99.15%, mean intersection over union (IoU) of 94.69%, mean accuracy of 97.73%, sensitivity of 96.02%, and F1 score of 94.88%, respectively. The plots of the loss and accuracy curve show that all the errors were small, with accuracy rates reaching 99.05%, 98.62%, 99.39%, and 98.97% for ASD, VSD, AVSD, and normal heart, respectively. The comparison accuracy of contour prediction for U-Net was 99.01%, while V-Net was 93.70%. This shows that the U-Net has better accuracy than the V-Net model architecture. It can be proven that the architecture of CNNs has been successful in segmenting the cardiac chamber to detect defects in the heart septum and support the work of cardiologists.
KW - Cardiac septal defect
KW - CNNs
KW - Contour segmentation
KW - Deep learning
UR - http://www.scopus.com/inward/record.url?scp=85106348104&partnerID=8YFLogxK
U2 - 10.1016/j.imu.2021.100601
DO - 10.1016/j.imu.2021.100601
M3 - Article
AN - SCOPUS:85106348104
VL - 24
JO - Informatics in Medicine Unlocked
JF - Informatics in Medicine Unlocked
SN - 2352-9148
M1 - 100601
ER -