A Deep Learning-Driven Pipeline for Differentiating Hypertrophic Cardiomyopathy from Cardiac Amyloidosis Using 2D Multi-View Echocardiography

Bo Peng, Xiaofeng Li, Xinyu Li, Zhenghan Wang, Hui Deng, Xiaoxian Luo, Lixue Yin, Hongmei Zhang

TL;DR

Distinguishing between Hypertrophic Cardiomyopathy (HCM) and Cardiac Amyloidosis (CA) is a notorious clinical challenge because both conditions present with similar heart wall thickening on ultrasounds. This paper introduces a multi-stage deep learning pipeline that integrates five different 2D echocardiographic views to achieve a 90.4% F1-score, significantly outperforming single-view and dual-view diagnostic methods.

What Problem Does This Paper Address?

In the world of cardiology, HCM and CA are "look-alike" diseases. HCM is a hereditary condition, while CA involves the deposition of insoluble proteins (amyloid) in heart tissue. Both lead to myocardial thickening and eventual heart failure, but their treatments are vastly different.

Currently, differentiating them via 2D echocardiography (ultrasound) is difficult even for experts, with misdiagnosis rates ranging from 10% to 30%. While advanced techniques like 2D speckle tracking (2D-STE) or biopsies exist, they are either computationally expensive, require high-end equipment, or are invasive. The authors sought to create an automated, high-accuracy pipeline using standard, widely available 2D multi-view echocardiography.

Methods

The researchers developed a three-stage "Deep Learning-Driven Pipeline":

  1. View Classification: The system first uses a Vision Transformer (ViT) to automatically sort ultrasound images into five standard views: Apical 4-chamber (A4C), Parasternal Long Axis (PLAX), and three Short Axis (PSAX) views (at the mitral valve, papillary muscle, and apex levels).
  2. Feature Extraction: A modified ResNet18 architecture extracts deep features from each of these five views independently. The authors chose ResNet18 because it is robust and less prone to overfitting on medical datasets compared to deeper models.
  3. Feature Fusion & Classification: The features from all five views are concatenated into a single vector. This "fused" representation is fed into a Linear Classifier to categorize the patient as having HCM, CA, or Normal cardiac function.

To ensure the model was actually looking at the heart, they used Grad-CAM, a visualization tool that generates heatmaps showing which parts of the image influenced the model's decision.

Datasets

The study utilized a private dataset from the Sichuan Provincial People’s Hospital (2018–2022):

  • View Classification: 13,575 images from 1,246 patients.
  • Disease Classification: Data from 242 patients (212 HCM and 30 CA) plus 200 healthy controls.
  • Data Augmentation: Because CA is relatively rare, the researchers used techniques like rotation, Gaussian blur, and contrast enhancement to expand the CA training set, ensuring the model didn't become biased toward the more common HCM.

Key Insights and Ideas

  • The Power of Multi-View: The most significant "aha moment" was the performance jump as more views were added. A single-view model (A4C) only achieved an F1-score of 0.704. Moving to two views (A4C + PLAX) raised it to 0.821. The full five-view model reached 0.904, proving that the structural nuances of these diseases are distributed across different cardiac perspectives.
  • Upstream Accuracy Matters: The authors performed an "error analysis" showing that if the initial stage misclassifies the view (e.g., mistaking one short-axis level for another), the final diagnostic accuracy plummets. This highlights that a robust pipeline is only as strong as its first step.
  • Transformers vs. CNNs: The paper strategically uses a Vision Transformer for view classification (where global context is key) but sticks to ResNet (CNN) for feature extraction, leveraging the strengths of both architectures.

Key Results

  • View Classification: Achieved 95% accuracy in identifying the correct echocardiographic view.
  • Disease Differentiation: The five-view model achieved a precision and recall of 0.905 and a micro-F1 score of 0.904.
  • Clinical Relevance: The model's performance is competitive with senior sonographers. Grad-CAM heatmaps confirmed the model focuses specifically on the ventricle and ventricular septum—the same areas human experts prioritize.
  • Classifier Comparison: While they tested SVMs and Random Forests, a simple linear network provided the best balance of accuracy and ease of integration into a deep learning pipeline.

Limitations and Future Work

The authors acknowledge a few hurdles before this can hit the clinic:

  • Sample Size: The number of CA patients (30) is small due to the rarity of the disease. While augmentation helped, a larger real-world sample is needed.
  • Single-Center Data: The model was trained on data from one hospital. Differences in ultrasound equipment and protocols at other institutions might affect performance.
  • Future Directions: The team suggests using Knowledge Distillation (training a smaller, faster model to mimic a larger one) to make the pipeline even more efficient for real-time use. They also propose Federated Learning as a way to train on multi-center data without compromising patient privacy.

Summary updated

Abstract

Hypertrophic cardiomyopathy (HCM) and cardiac amyloidosis (CA) are both heart conditions that can progress to heart failure if untreated. They exhibit similar echocardiographic characteristics, often leading to diagnostic challenges. This paper introduces a novel multi-view deep learning approach that utilizes 2D echocardiography for differentiating between HCM and CA. The method begins by classifying 2D echocardiography data into five distinct echocardiographic views: apical 4-chamber, parasternal long axis of left ventricle, parasternal short axis at levels of the mitral valve, papillary muscle, and apex. It then extracts features of each view separately and combines five features for disease classification. A total of 212 patients diagnosed with HCM, and 30 patients diagnosed with CA, along with 200 individuals with normal cardiac function(Normal), were enrolled in this study from 2018 to 2022. This approach achieved a precision, recall of 0.905, and micro-F1 score of 0.904, demonstrating its effectiveness in accurately identifying HCM and CA using a multi-view analysis.
BibTeX
@misc{peng2024deep,
  author = {Bo Peng and Xiaofeng Li and Xinyu Li and Zhenghan Wang and Hui Deng and Xiaoxian Luo and Lixue Yin and Hongmei Zhang},
  title = {{A Deep Learning-Driven Pipeline for Differentiating Hypertrophic Cardiomyopathy from Cardiac Amyloidosis Using 2D Multi-View Echocardiography}},
  year = {2024},
  eprint = {2404.16522},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2404.16522},
}
Figure