CardiacField: Computational Echocardiography for Universal Screening

Chengkang Shen1, Hao Zhu1, You Zhou1, Yu Liu2, Si Yi1, Lili Dong2, Weipeng Zhao2, David J. Brady3, Xun Cao1, Zhan Ma1, Yi Lin2
1Nanjing University, 2Fudan University, 3University of Arizona
Paper Code

Abstract

Cardiovascular diseases, the worldwide leading cause of death, are preventable and treatable. Early diagnosis and monitoring using ultrasound, x-ray or MRI are crucial clinical tools. Routine imaging is, however, currently cost prohibitive. Here we demonstrate the use of computational imaging to reduce the cost of tomographic echocardiography by >1000 × while also improving image quality and diagnostic utility. This advance relies on decompressive inference using artificial neural networks. Our system, CardiacField, utilizes 2DE probes to provide accurate and automated assessments of cardiac function, eliminating the need for specialized professional training. CardiacField generates a 3D heart from 2D echocardiograms with <2 minute processing time. The system automatically segments and quantifies the volume of the left ventricle (LV) and right ventricle (RV) without manual calibration. CardiacField estimates the left ventricular ejection fraction (LVEF) with 48% higher accuracy than state-of-the-art video-based methods, and the right ventricular ejection fraction (RVEF) with a similar accuracy, which is not available in existing 2DE methods. This technology will enable routine world-wide tomographic heart screening, such that patients will get instant feedback on lifestyle changes that improve heart health. CardiacField also illustrates the value of a conceptual shift in diagnostic imaging from direct physical model inversion to Bayesian inference.

The workflow of the CardiacField

Description of the image
The workflow of the CardiacField and EF calculation. (a) The whole 3D heart is represented as an implicit function, where the input is a 3D coordinate of the heart and the output is the corresponding intensity. This continuous 3D function is approximated by an MLP network with multiresolution hash-table. The implicit function is determined by minimizing a physical-informed loss function. (b) 2D echocardiographic images are acquired by rotating the 2DE probe in 360 degrees around the apex of the heart. Then we synchronize multiple cardiac views and select images at end-diastole and end-systole based on concurrently recorded ECG. (c) The 3D rendering of the reconstructed heart by the CardiacField. (d) The CardiacField represents a 3D heart in a continuous implicit function, leading to less artifacts in slices compared with the conventional interpolation as indicated by the yellow arrows. (e) The workflow of EF calculation based on the CardiacField. We first perform the uniform sampling on the reconstructed 3D heart to generate about 20-30, 3mm-thick 2D slices parallel to the apical four-chamber view, and then use the segmentation model developed in EchoNet to classify the LV and RV regions. We calculate the volumes of the LV and RV by summing the area of each slice. The EF is defined as the ratio of changes in the ESV and EDV of LV/RV. (MLP: multilayer perceptron. ECG: electrocardiogram. LV: left ventricular. RV: right ventricular. ESV: end-systolic volume. EDV: end-diastolic volume.)

Results

Volume segmentation and analysis of LVEF and RVEF

Description of the image
(a) We use the CardiacField to reconstruct realistic 3D heart with real-captured 2D echocardiographic images. (b) We compare the same cross-sectional views (apical four-chamber view) between the reconstructed hearts by CardiacField and the real-captured ones by the 3D probe for 10 independent patients. The 3D cardiac volumes reconstructed by Our CardiacField exhibit more spatial details and less artifacts than those rendered by the 3DE probe. (c) We compare our LVEF results with EchoNet, where the EFs obtained by the 3D ultrasound machine (after calibration by the experienced sonographers) are set as the ground truth. The red and blue lines represent the least squares regression line between model prediction and ground truth, respectively. Our method provides more accurate LVEF prediction with over 48% improvement. (d) We also calculate the RVEF results, which is not available in general for 2DE-based methods like EchoNet. (e) We reconstruct the 3D hearts and calculate the EFs for another 5 volunteers using 2D images from two different 2D ultrasound machines widely used in clinics. We can achieve stable results for different machines, which demonstrate the generalization ability of our method.

3D dynamic heart within one cardiac cycle

Description of the image
We show some snapshots of a reconstructed 3D dynamic heart within one cardiac cycle using the CardiacField, compared with that acquired by the 3DE probe. The ECG is used to indicate the phases within one cardiac cycle.

Analysis of 3D reconstruction

Description of the image
(a) Visualization of the positional parameter trajectories. The initial and refined positional parameters are obtained using PlaneInVol and CardiacField, respectively. (b) We compare the real-captured 3D heart (as ground truth) with the reconstructed 3D hearts by the CardiacField and the conventional interpolation method. The CardiacField and interpolation method use 'Input Views' for 3D heart reconstruction. For PSNR, higher values indicate better performance. 'Long-Axis', 'Short-Axis' and '3D Volume' denote the evaluation scores for newly generated cardiac long-axis views, newly generated cardiac short-axis views and reconstructed 3D heart, respectively. (c) Illustration of the 'continuous-slicing' functionality of our CardiacField compared to the conventional interpolation method. The same views are extracted for different methods. From top to bottom, the cross-sectional views of real-captured 3D heart, the interpolation method (without position refinement) and our CardiacField. The CardiacField is able to generate arbitrary views while avoiding interpolation artifacts highlighted by yellow arrows.

Volume Segmentation Results

BibTeX

@article{shen2023cardiacfield,
  author    = {Shen, Chengkang and Zhu, Hao and Zhou, You and Liu, Yu and Yi, Si and Dong, Lili and Zhao, Weipeng and Brady, David and Cao, Xun and Ma, Zhan and Lin, Yi},
  title     = {CardiacField: Computational Echocardiography for Universal Screening},
  journal   = {Research Square},
  year      = {2023},
}