Direct Multitype Cardiac Indices Estimation via Joint Representation and Regression Learning

Wufeng Xue, Ali Islam, Mousumi Bhaduri, Shuo Li

Introduction

In clinical cardiology, accurately measuring cardiac indices—such as myocardial wall thickness and ventricular cavity area—is essential for diagnosing and monitoring heart disease. Traditionally, these measurements rely on a segmentation-first approach, where specialists or automated tools trace the boundaries of heart structures before calculating geometric values. However, this process is computationally intensive and sensitive to image quality and the heart's complex motion.

Comparison of the proposed joint learning method against existing two-phase and traditional segmentation-based approaches.

While "direct methods" exist to estimate values without explicit segmentation, they typically follow a rigid two-phase pipeline: feature extraction followed by a separate regression model. Because these steps are trained independently, the extracted features may not be optimized for the specific clinical measurements required.

Methodology: Indices-Net

To address these limitations, researchers developed Indices-Net, a framework that estimates multiple cardiac health indicators directly from MR images. The system employs a Joint Learning strategy, where a representation model and a regression model are trained simultaneously.

The architecture consists of two tightly coupled components:

  1. Deep Convolutional Autoencoder (DCAE): This component learns a compressed representation of the cardiac image. It includes an encoder to extract features and a decoder to reconstruct the image, ensuring the network captures the underlying anatomy.
  2. Multiple Output CNN: This regressor takes the features from the DCAE to predict eight distinct cardiac indices: six regional wall thicknesses (WT1–WT6) and two planar areas (A1 and A2).
Visualization of the eight cardiac indices estimated by Indices-Net: six regional wall thicknesses (WT1-WT6) and two planar areas (A1, A2).

Unlike two-phase methods, Indices-Net back-propagates the error from the predicted indices through both the regressor and the autoencoder. This forces the autoencoder to refine its internal representation to be specifically "indices-relevant." Additionally, a pre-training step called Image-DCAE is used to help the model learn general cardiac structures from unlabeled images before focusing on the regression task.

Experimental Results

The model was evaluated using a dataset of 2,900 short-axis cine MR images from 145 subjects, encompassing various pathologies and data from multiple hospitals and scanner vendors. Indices-Net demonstrated significant improvements over traditional methods:

  • High Accuracy: The framework achieved a Mean Absolute Error (MAE) of just 1.44 ± 0.71 mm for wall thickness and 204 ± 133 mm² for areas.
  • Superior Performance: Compared to traditional segmentation, it reduced measurement error by 55.1% for wall thickness and 17.4% for areas.
  • Temporal Consistency: Unlike segmentation-based methods that struggle with heart motion, Indices-Net maintained consistently low error across all 20 frames of the cardiac cycle.
  • Clinical Utility: The model effectively calculated functional metrics like Ejection Fraction (EF), a primary indicator of heart health.

Discussion and Future Work

A key insight from the study is that joint learning allows the network to focus on specific anatomical regions for different metrics—for instance, focusing on the epicardial boundary for wall thickness while looking at the blood pool for cavity area. This "task-aware" focus even helps the model infer measurements in regions where boundaries are typically "invisible" or difficult for human tracers to define.

While the current workflow requires manual landmarking for region-of-interest cropping, future iterations aim to automate this step for a fully end-to-end system. Additionally, extending the framework to 3D volumetric data or incorporating temporal modeling via Recurrent Neural Networks (RNNs) represents the next logical step in this research.

Summary updated

BibTeX
@article{xue2017direct,
  author = {Wufeng Xue and Ali Islam and Mousumi Bhaduri and Shuo Li},
  title = {{Direct Multitype Cardiac Indices Estimation via Joint Representation and Regression Learning}},
  journal = {IEEE TMI},
  year = {2017},
  doi = {10.1109/tmi.2017.2709251},
}
Figure