Machine-Learning Tools

Our group develops machine-learning approaches that automate the analysis of cardiac images. Tools trained using expert annotations can be applied to large datasets, reduce variability, and make advanced analyses such as regional strain practical in routine care.

We have also shown that how image data is represented affects the accuracy of these tools.

Automated Regional Right Ventricular Function

Measuring regional right ventricular (RV) function from cineCT previously required time-intensive, semi-automated segmentation of the RV blood pool. We trained a deep learning right heart segmenter on cineCT from patients with normal function, heart failure, pulmonary hypertension, and adult congenital heart disease.

Combined with our regional strain analysis, the fully automated pipeline accurately reproduces RV volumes and free-wall and septal strain in an independent test cohort.

Craine et al., European Heart Journal - Imaging Methods and Practice 2026, Figure 1

Recent publications include:


Preprints & conference abstracts:

Automated CT-based regional RV strain assessment is successful with U-Net methods. Craine A, Severance L, Alshawabkeh L, Kim N, Adler E, Narezkina A, Ben-Yehuda O, Contijoch F. SCCT 2025 Annual Scientific Meeting (poster)

Intravascular Ultrasound Segmentation

Cardiac allograft vasculopathy is a leading cause of late graft failure after heart transplantation and is monitored with intravascular ultrasound (IVUS). We are developing automated whole-vessel IVUS segmentation and methods to evaluate segmentation quality.

Preprints & conference abstracts:

Whole vessel automated segmentation of IVUS images for evaluation of cardiac allograft vasculopathy. Lyu N, et al. ISHLT 45th Annual Meeting, 2025

Quantitative evaluation of IVUS image segmentations. Lyu N, et al. Biomedical Engineering Society Annual Meeting, 2024

Detecting Wall Motion Abnormalities

Regional wall motion abnormalities are a key sign of coronary disease and cardiomyopathy. We have automated their detection from 4D cardiac CT in several ways: deep learning applied to volume-rendered videos of the left ventricle, automated longitudinal strain measured in long-axis views, and thresholds on regional shortening derived from our SQUEEZ analysis.

Chen et al., Frontiers in Cardiovascular Medicine 2022, Figure 1

Li et al., Frontiers in Cardiovascular Medicine 2022, Figure 2

Memory Efficient Image Representation Improves Boundary Segmentation

CT image volumes are large and consume much of the memory available on a GPU, which limits the complexity of the models that can be applied. Octrees represent images in a memory-efficient way, and an intensity threshold controls how much the image is compressed.

Trading a small amount of image fidelity for greater model complexity improved segmentation of the left atrium and left ventricle, particularly at chamber boundaries.

Automated Quantification of Function and Identification of Short- and Long-axis Imaging Planes

We developed a convolutional neural network (CNN)-based method that both quantifies heart size and function and identifies the 2D imaging planes cardiologists conventionally use to visualize the heart. A shared architecture with transfer learning between the labeling and slice-planning tasks improved training with limited annotated data.

The automated planes agree closely with those chosen by experts, and the framework makes it easy to compare patients over time. The example below shows a patient at baseline and at 3-month follow-up.

Short- and long-axis slices at baseline