Automatic Morphological Analysis of the Left Ventricle

Context

This project explored whether cardiac shape changes could be described from echocardiography video with discrete morphological descriptors instead of relying only on post-mortem analysis or manual clinical interpretation.

The original paper, available as ISICS.pdf, was written in Spanish with Nidiyare Hevia-Montiel and Ana Paulina Haro Alvarez. It focused on the left ventricle and interventricular septum in a murine model, using echocardiographic sequences as the input signal.

The motivation was straightforward: cardiac function is dynamic, but many structural and molecular insights still depend on invasive or post-mortem workflows. If shape descriptors can be extracted over time from medical images, they can help describe ventricular behavior during a cardiac sequence and support future longitudinal analysis.

Research Goal

The goal was to analyze temporal changes in the morphology of two regions of interest:

• The left ventricle cavity.
• The interventricular septum.

The work tested whether two-dimensional contour descriptors could capture differences and similarities across video frames. The study looked especially at descriptors derived from slope-chain codes, including tortuosity, non-circularity, compactness, perimeter, area, and a vector-based morphological signature.

This was an early-stage research prototype, so the success criterion was not clinical diagnosis. The target was a repeatable processing pipeline that could turn ultrasound video into contour data and then into measurable shape indices.

Data Acquisition

The study used echocardiography from two male ICR mice. The animals were handled under the applicable laboratory animal-care protocol, anesthetized, prepared for imaging, and scanned with a high-frequency linear transducer.

The acquisition captured B-mode cardiac images in long-axis and short-axis views. For the long-axis five-chamber view, the transducer was placed along the left hemithorax. The resulting images and videos were stored for later processing.

The research setup mattered because the input was not a clean synthetic shape. It was ultrasound video: noisy, frame-based, and dependent on acquisition angle, biological motion, and image resolution.

Video Processing Pipeline

The first technical step was converting echocardiography video into a format suitable for volumetric analysis.

Each frame from the video sequence was loaded into a four-dimensional matrix containing spatial coordinates, color information, and time. Because the analysis software worked primarily with three-dimensional NIfTI images, the video was converted to grayscale and transformed so time could be treated as a volume axis.

That conversion produced a stack of cardiac frames that could be inspected and segmented in BrainVisa-Anatomist. In practical terms, the video became a navigable image volume where each slice represented a point in time.

Region Segmentation

For this preliminary study, segmentation was manual.

Every third frame was inspected, and the regions of interest were delineated by hand:

• The left ventricle cavity.
• The interventricular septum.

After the masks were created, a Moore-neighborhood border detection algorithm extracted discrete contours from each binary region. Those contours became the input to the descriptor pipeline.

Manual segmentation was a reasonable choice at this stage because the research question was about descriptor behavior, not automated ultrasound segmentation. It kept the early prototype focused on whether the shape measurements were useful once a contour existed.

Discrete Shape Descriptors

The core method used slope-chain codes. A contour is represented by walking around the curve with fixed-length line segments and measuring the slope changes between consecutive segments. These changes are scaled into a normalized range, which makes the representation independent of translation and rotation.

From that chain representation, the study computed scalar and vector descriptors.

Scalar descriptors summarize the whole contour. Examples include accumulated slope change, average slope change, tortuosity, non-circularity, compactness, perimeter, and area. These values make it possible to plot a region’s morphology over time and compare trends across frames.

Vector descriptors preserve more local shape information. The morphological signature used in the paper stores the distance between contour elements and the centroid of the closed curve. Unlike some of the scalar descriptors, this signature can retain enough information to reason about the contour shape more directly, though it is still affected by scale.

Results

The preliminary analysis plotted scalar descriptors across 42 sampled frames from the echocardiography videos.

The most important descriptor discussed in the results was non-circularity. For each case study, non-circularity values were plotted for the left ventricle and the interventricular septum. The resulting curves showed correspondence between the ventricle and septum trends within the same animal.

The analysis also compared the same region across the two animals. The left ventricle curves appeared to have similar behavior, but the study was careful about this point: without synchronizing frames to a physiological cardiac reference, direct inter-subject comparison is limited.

The project also selected frames with minimum and maximum non-circularity values and overlaid the corresponding left ventricle regions. This visual check helped verify that the numeric descriptor reflected meaningful contour differences rather than only abstract calculation.

Key Limitation

The main limitation was synchronization.

A video frame number is not the same thing as a physiological cardiac phase. Without aligning frames to events such as systole and diastole, descriptor curves can show trends, but they cannot yet support strong comparisons between subjects.

That limitation shaped the conclusion of the work. The descriptor pipeline was promising, but a stronger protocol would need cardiac-cycle registration, image-resolution analysis, and a clearer relationship between the computed descriptors and physiological events.

What I Took From It

This project was an early example of a pattern I still value: take a messy real-world signal, reduce it into a structured representation, and keep the assumptions visible.

The pipeline had clear boundaries:

• Acquisition produced echocardiography video.
• Conversion turned video into a frame volume.
• Segmentation produced binary masks.
• Border detection produced discrete contours.
• Descriptor extraction produced scalar and vector measurements.
• Plotting and overlays helped validate whether the measurements matched visible shape changes.

That separation made the prototype easier to reason about. Each stage had its own uncertainty, and each output could be inspected before moving to the next step.

Outcome

The final result was a preliminary research workflow for describing left ventricle and interventricular septum morphology from murine echocardiography.

It did not claim a diagnostic model. Instead, it showed that discrete contour descriptors such as non-circularity, tortuosity, and morphological signatures could be extracted over time from ultrasound-derived cardiac regions.

The most useful outcome was the next question it made obvious: to make these descriptors clinically or experimentally comparable, the image analysis pipeline needs to be synchronized with cardiac physiology. Once frame timing, cardiac phase, and acquisition variability are controlled, the same descriptor framework can become a stronger tool for longitudinal cardiac-shape analysis.