[Paper Review] Segmentation of points of interest during fetal cardiac assesment in the first trimester from color Doppler ultrasound

Ruxandra Stoean, Dominic Gabriel Iliescu+1 more|arXiv (Cornell University)|Sep 26, 2019

Phonocardiography and Auscultation Techniques14 citations

TL;DR

This study proposes a traditional machine learning approach using Zernike moments and distance-based classification to automatically detect key fetal cardiac views—V, X, and parallel red flows—in first-trimester color Doppler ultrasound. Despite a small dataset, the method achieved promising accuracy in identifying clinically relevant frames, supporting early screening for congenital heart disease.

ABSTRACT

The present paper puts forward an incipient study that uses a traditional segmentation method based on Zernike moments for extracting significant features from frames of fetal echocardiograms from first trimester color Doppler examinations. A distance based approach is then used on the obtained indicators to classify frames of three given categories that should be present in a normal heart condition. The computational tool shows promise in supporting the obstetrician in a rapid recognition of heart views during screening.

Motivation & Objective

• To develop a computational tool that supports obstetricians in rapidly identifying critical fetal cardiac views during first-trimester screening.
• To address the challenge of low detection rates for congenital heart disease (CHD) in the first trimester due to underdeveloped heart structures.
• To classify ultrasound frames into three clinically relevant categories: V-shaped blue flow, X-shaped blue flow, and parallel red flows.
• To explore the feasibility of using traditional image features (Zernike moments) for automated segmentation and classification in low-resource, early-stage fetal echocardiography.

Proposed method

• Extracted frames from first-trimester color Doppler ultrasound videos and separated them into blue (V, X, other) and red (parallel, other) classification tasks.
• Applied binary segmentation to isolate regions of interest (ROIs) for each color, focusing on distinct flow patterns.
• Computed 25 Zernike moments per segmented ROI, combined with image dimensions and polygon count, to form a 28-feature vector per frame.
• Used a distance-based classification (nearest neighbor) to assign test frames to one of the three classes, due to limited training data.
• Ensured data diversity by training on frames from different patients than those used in testing, enhancing generalization potential.
• Retained morphological and topological features (e.g., area, perimeter, Delaunay triangulation) as potential future enhancements.

Experimental results

Research questions

• RQ1 Can Zernike moments effectively extract discriminative features from first-trimester fetal color Doppler ultrasound frames for cardiac view classification?
• RQ2 To what extent can a distance-based classifier achieve reliable classification of V, X, and parallel red flow patterns with limited training data?
• RQ3 Can automated detection of key hemodynamic signatures (V, X, parallel red) reduce the workload of sonographers during early fetal cardiac screening?
• RQ4 How do variations in image appearance and low contrast affect the performance of traditional feature-based methods in early fetal echocardiography?

Key findings

• The method achieved high accuracy in identifying V and X patterns in blue flow, with physicians confirming most detections as correct.
• False positives occurred primarily when 'other' class frames were misclassified as V, indicating challenges in distinguishing subtle non-specific patterns.
• For red flow classification, the parallel line pattern was mostly correctly detected, though some 'other' frames were incorrectly labeled as parallel.
• The small training set (53 frames: 20 V, 12 X, 21 other) still yielded promising results, suggesting feasibility for clinical support despite data scarcity.
• The distance-based classifier performed adequately with limited data, demonstrating proof of concept for feature-based automation in early fetal cardiac assessment.
• The approach significantly reduces the number of irrelevant frames a sonographer must manually review, potentially accelerating learning for residents and improving screening efficiency.