Diagnosing Cardiac Deformations using 3d Optical Flow

Vithu Logan Jeya, John ZelekVision 2015 · DOI

Abstract

This paper explores the viability of applying 3D optical flow techniques on 3D heart sequences to diagnose cardiac abnormalities and disease. Tagged magnetic resonance imaging (TMRI) is a non-invasive method to visualize in vivo myocardium motion during a cardiac cycle. By tracking the 3D trajectories of tagged material points, it is possible to construct a volumetric model of the heart. Specifically, we use generated meshless deformable models (MDM) which describe an object as a point cloud inside the object boundary. We extend the 2D least squares and regularization approaches of Lucas and Kanade to 3D in order to capture the flow, specifically the contraction and expansion of various parts of the heart motion. Features are extracted from this flow and a rudimentary SVM is used to classify unhealthy hearts.

Introduction

Cardiovascular disease (CVD) remains the leading cause of death globally. While early detection significantly improves survival rates, diagnosing subtle abnormalities in heart wall motion is notoriously difficult. The heart doesn't just pump; it twists, expands, and contracts in a complex 3D pattern. Standard 2D imaging often fails to capture the full volumetric "wringing" motion of the left ventricle, which is a primary indicator of cardiac health. While Tagged Magnetic Resonance Imaging (TMRI) allows clinicians to visualize this motion non-invasively, there is a lack of automated, robust tools to translate these high-resolution 3D images into reliable diagnostic classifications.

Methods

The proposed pipeline moves from raw 3D imagery to a binary "healthy/unhealthy" diagnosis using three main stages:

  1. 3D Lucas-Kanade Optical Flow: The researchers extended the classic 2D Lucas-Kanade algorithm to three dimensions. By solving a 3D motion constraint equation ($I_x U + I_y V + I_z W = -I_t$), they calculate velocity vectors for a "point cloud" representation of the heart, known as a Meshless Deformable Model (MDM).
  2. Motion Descriptors: To make sense of the massive amount of flow data, they created per-frame orientation histograms (similar to Histogram of Oriented Gradients, or HOG, used in object detection). They then extracted six temporal statistics from these histograms—including Maximum, Mean, and Standard Deviation—to describe how the heart's motion evolves over a full cardiac cycle.
  3. SVM Classification: These temporal features were fed into a Support Vector Machine (SVM) using a Radial Basis Function (RBF) kernel. This "rudimentary" classifier was tasked with the binary job of separating healthy hearts from those with abnormalities.
Visualization of 3D optical flow vectors (blue arrows) on the heart model.

Datasets and Experimental Setup

The study utilized a specialized dataset provided by the Shenzhen Institute of Advanced Technology, consisting of 20 heart sequences (10 healthy, 10 unhealthy). The data type used was Tagged MRI (TMRI) converted into 3D meshless deformable models. An experimental split of 70/30 was used, with 14 patients for training and 6 for testing.

Key Results

Despite the small sample size, the preliminary results are promising for a proof-of-concept. The model achieved an overall accuracy of 83%. Notably, it demonstrated 100% specificity, meaning it was perfect at identifying healthy hearts. The sensitivity was lower at 66%, indicating the model struggled more with "false negatives," or missing some unhealthy cases.

Performance metrics for the SVM classifier on the heart dataset.

The high specificity suggests that this approach could be particularly useful as a screening tool to "rule out" healthy patients with high confidence.

Discussion and Future Work

By treating the heart as a 3D volume changing over time (4D), the model captures physiological nuances—like the specific contraction and expansion phases—that are often lost in 2D slices. The use of quantized 3D principal directions helps the model focus on the intent of the motion, making it more robust to noise in the MRI data.

However, the authors acknowledge several limitations. With only 20 patients, the results are an evaluation of potential rather than a definitive clinical measure. Larger datasets are required to validate the SVM's performance. Additionally, MRI data can be discontinuous between frames; future work should focus on mitigating these "jumps" to improve the accuracy of the optical flow calculation. Finally, there is a need to establish a standardized, large-scale 4D heart dataset to allow researchers to benchmark different motion-tracking algorithms more effectively.

Summary updated

Abstract

This paper explores the viability of applying 3D optical flow techniques on 3D heart sequences to diagnose cardiac abnormalities and disease. Tagged magnetic resonance imaging (TMRI) is a non-invasive method to visualize in vivo myocardium motion during a cardiac cycle. By tracking the 3D trajectories of tagged material points it is possible to construct a volumetric model of the heart. Specifically, we use generated meshless deformable models (MDM) which describe an object as a point cloud inside the object boundary. We extend the 2D least squares and regularization approaches of Lucas and Kanade to 3D in order capture the flow, specifically the contraction and expansion of various parts of the heart motion. Features are extracted from this flow and a rudimentary SVM is used to classify unhealthy hearts.
BibTeX
@inproceedings{jeya2015diagnosing,
  author = {Vithu Logan Jeya and John Zelek},
  title = {{Diagnosing Cardiac Deformations using 3d Optical Flow}},
  booktitle = {Vision},
  year = {2015},
  doi = {10.15353/vsnl.v1i1.55},
}
Figure