FetalNet: Low-light fetal echocardiography enhancement and dense convolutional network classifier for improving heart defect prediction. (2022)

• Record Type: Journal Article
• Authors: Sutarno, Sutarno; Nurmaini, Siti; Partan, Radiyati Umi; Sapitri, Ade Iriani; Tutuko, Bambang; Naufal Rachmatullah, Muhammad; Darmawahyuni, Annisa; Firdaus, Firdaus; Bernolian, Nuswil; Sulistiyo, Deny
• Is Part Of: Informatics in medicine unlocked. Volume 35(2023)
• Publication Date: 2022
• Subjects: Deep learning -- Ultrasound -- Congenital heart disease -- Low light image enhancement; Medical informatics -- Periodicals; 610.285
• DOI: 10.1016/j.imu.2022.101136
• Languages: English
• ISSNs: 2352-9148

Abstract

Abstract: Background: Fetal heart defect (FHD) examination by ultrasound (US) is challenging because it involves low light, contrast, and brightness. Inadequate US images of fetal echocardiography play an important role in the failure to detect FHDs manually. The automatic interpretation of fetal echocardiography was proposed in a previous study. However, the low quality of US images reduces the prediction rate of computer-assisted diagnosis results. Methods: To increase the FHD prediction rate, we propose low-light fetal echocardiography enhancement stacking with a dense convolutional network classifier named "FetalNet." Our proposed FetalNet model was developed using 460 US images to produce an image enhancement model. The results showed that all raw US images could be improved with satisfactory performance in terms of increasing the peak signal-to-noise ratio of 30.85 dB, a structural similarity index of 0.96, and a mean squared error of 18.16. Furthermore, all reconstructed images were used as inputs in a convolutional neural network to generate the best classifier for predicting FHD. Results: The proposed FetalNet model increased the FHD prediction rate by approximately 25% in terms of accuracy, sensitivity, and specificity and produced 100% predictive negative using unseen data. Conclusions: The proposed deep learning model has the potential to identify FHD accurately and shows potential for practical use in identifying congenital heart diseases in the future.