Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Congenital Heart Disease recognition using Deep Learning/Transformer models

Aidar Amangeldi, Vladislav Yarovenko, Angsar Taigonyrov

TL;DR

Addressing the challenge of non-invasive CHD detection and false negatives in newborn screening, the paper investigates a dual-modality deep-learning framework that leverages audio (heart sounds) and chest X-ray images. It evaluates multiple audio representations (STFT, Mel, GAF) with CNN backbones and applies preprocessing and transformer-based image classifiers on DICOM-derived X-ray data, followed by late-fusion strategies to combine predictions. The strongest results are achieved with STFT + ResNet-50V2 for audio and ResNet-18 for DICOM-XRAY, achieving 71% and 80.72% accuracy, respectively; preprocessing steps boosted DICOM performance by 6-7%. The findings highlight the potential of combining acoustic and imaging modalities for CHD screening, while also underscoring the trade-offs between model complexity, data requirements, and practical deployability in resource-limited settings.

Abstract

Congenital Heart Disease (CHD) remains a leading cause of infant morbidity and mortality, yet non-invasive screening methods often yield false negatives. Deep learning models, with their ability to automatically extract features, can assist doctors in detecting CHD more effectively. In this work, we investigate the use of dual-modality (sound and image) deep learning methods for CHD diagnosis. We achieve 73.9% accuracy on the ZCHSound dataset and 80.72% accuracy on the DICOM Chest X-ray dataset.

Congenital Heart Disease recognition using Deep Learning/Transformer models

TL;DR

Addressing the challenge of non-invasive CHD detection and false negatives in newborn screening, the paper investigates a dual-modality deep-learning framework that leverages audio (heart sounds) and chest X-ray images. It evaluates multiple audio representations (STFT, Mel, GAF) with CNN backbones and applies preprocessing and transformer-based image classifiers on DICOM-derived X-ray data, followed by late-fusion strategies to combine predictions. The strongest results are achieved with STFT + ResNet-50V2 for audio and ResNet-18 for DICOM-XRAY, achieving 71% and 80.72% accuracy, respectively; preprocessing steps boosted DICOM performance by 6-7%. The findings highlight the potential of combining acoustic and imaging modalities for CHD screening, while also underscoring the trade-offs between model complexity, data requirements, and practical deployability in resource-limited settings.

Abstract

Congenital Heart Disease (CHD) remains a leading cause of infant morbidity and mortality, yet non-invasive screening methods often yield false negatives. Deep learning models, with their ability to automatically extract features, can assist doctors in detecting CHD more effectively. In this work, we investigate the use of dual-modality (sound and image) deep learning methods for CHD diagnosis. We achieve 73.9% accuracy on the ZCHSound dataset and 80.72% accuracy on the DICOM Chest X-ray dataset.
Paper Structure (15 sections, 2 equations, 6 figures, 1 table)

This paper contains 15 sections, 2 equations, 6 figures, 1 table.

Table of Contents

  1. Introduction
  2. Related Works
  3. Spectrogram based approaches
  4. X-ray based approaches
  5. Dataset description
  6. ZCHSound
  7. DICOM
  8. Methodology
  9. Audio Late Fusion
  10. Classifiication models for DICOM CHD-CXR images dataset
  11. Results and Discussion
  12. Audio Late Fusion
  13. DICOM image classification
  14. Conclusion
  15. References

Figures (6)

  • Figure 1: DICOM: ASD, Normal, PDA and VSD types of CHD.
  • Figure 2: STFT (a), Mel (b), and GAF (c) Transformations of Audio Data.
  • Figure 3: An Overview of the Audio Processing Methodology.
  • Figure 4: An Overview of the Image Processing Methodology.
  • Figure 5: Test Accuracy Results of Selected Base Model and Late Fusion Techniques.
  • ...and 1 more figures