BMC Bioinformatics (Sep 2023)

Automatic echocardiographic anomalies interpretation using a stacked residual-dense network model

  • Siti Nurmaini,
  • Ade Iriani Sapitri,
  • Bambang Tutuko,
  • Muhammad Naufal Rachmatullah,
  • Dian Palupi Rini,
  • Annisa Darmawahyuni,
  • Firdaus Firdaus,
  • Satria Mandala,
  • Ria Nova,
  • Nuswil Bernolian

DOI
https://doi.org/10.1186/s12859-023-05493-9
Journal volume & issue
Vol. 24, no. 1
pp. 1 – 21

Abstract

Read online

Abstract Echocardiographic interpretation during the prenatal or postnatal period is important for diagnosing cardiac septal abnormalities. However, manual interpretation can be time consuming and subject to human error. Automatic segmentation of echocardiogram can support cardiologists in making an initial interpretation. However, such a process does not always provide straightforward information to make a complete interpretation. The segmentation process only identifies the region of cardiac septal abnormality, whereas complete interpretation should determine based on the position of defect. In this study, we proposed a stacked residual-dense network model to segment the entire region of cardiac and classifying their defect positions to generate automatic echocardiographic interpretation. We proposed the generalization model with incorporated two modalities: prenatal and postnatal echocardiography. To further evaluate the effectiveness of our model, its performance was verified by five cardiologists. We develop a pipeline process using 1345 echocardiograms for training data and 181 echocardiograms for unseen data from prospective patients acquired during standard clinical practice at Muhammad Hoesin General Hospital in Indonesia. As a result, the proposed model produced of 58.17% intersection over union (IoU), 75.75% dice similarity coefficient (DSC), and 76.36% mean average precision (mAP) for the validation data. Using unseen data, we achieved 42.39% IoU, 55.72% DSC, and 51.04% mAP. Further, the classification of defect positions using unseen data had approximately 92.27% accuracy, 94.33% specificity, and 92.05% sensitivity. Finally, our proposed model is validated with human expert with varying Kappa value. On average, these results hold promise of increasing suitability in clinical practice as a supporting diagnostic tool for establishing the diagnosis.

Keywords