Comparison of Computer Vision Approaches and Machine Learning Integration in Cardiac Kinematics Analysis

By Samudrapom DamReviewed by Susha Cheriyedath, M.Sc.Jan 31 2024

In an article recently published in the journal Scientific Reports, researchers performed a comparative analysis of three computer vision approaches to evaluate contraction kinematics in both ventricular isolated single cells and cardioids and used machine learning (ML) algorithms to test the prediction performance of the three training datasets generated from each approach.

Background

The beating heart is primarily the outcome of optimal electromechanical synchronization. Researchers have developed several models to quantify and unveil mechanical and electrical activities both from whole-organ to cellular levels owing to the high interconnection between them.

Recent improvements in computation ability have enabled the evaluation of both heart functions/electrical and mechanical using in-silico simulations in place of animal experimentation. In-vitro studies experience several challenges compared to in-vivo and ex-vivo studies while measuring the heart's mechanical activity. Specifically, the contraction force and kinematics measured in vitro require addressing the micrometric cell forces expressed and dimensions. This also necessitates expensive and specific instrumentation.

In biology, the emergence of computer vision technologies has introduced alternative solutions for cell movement analysis and tracking. Specifically, motion tracking analysis using high-temporal and spatial resolution cameras can record the entire kinematics evaluation. Several algorithms based on information obtained from video frames have been implemented to develop user-friendly, highly flexible, and open-source software to estimate cellular kinematics from video recordings.

For instance, MUSCLEMOTION is one of the most commonly utilized open-source software based on image intensity-based tracking and segmentation. Similarly, CONTRACTIONWAVE has been validated as an effective tracking algorithm in a recent study. However, the performance of these software varies based on the analyzed specimens/samples.

The proposed approach

In this study, researchers compared the performances of different software in cardioids, in-silico models, and in-vitro mouse adult ventricular cardiomyocytes. The study's objective was to evaluate contraction kinematics in both cardioids and ventricular isolated single cells using different approaches and compare their performance.

In-vitro high-resolution videos were acquired at 0.5-1-2 Hz during suprathreshold stimulation. Moreover, the tested samples  were also exposed to depolarizing and inotropic substances. The in silico and in vitro videos were analyzed/kinematic analyses were performed using three computer vision methods, including CONTRACTIONWAVE tracking software, MUSCLEMOTION open-source software, and ViKiE, which is customized in-house video kinematic evaluation software.

MUSCLEMOTION was used as the benchmark. ViKiE is primarily a pipeline containing an open-source tracking software/Video Spot Tracker (version 08.11) and a custom MATLAB^® script utilized in motion analysis. The Video Spot Tracker tracks the sample motion using a proper marker as in feature-based tracking systems.

Three ML algorithms were used to evaluate the robustness of the motion-tracking approaches/the prediction performance of three training datasets created from every software and their overall sensibility in the eventual kinematic changes after the administration of inotropic substances. Specifically, researchers explored the potential of ML to improve the accuracy of movement analysis.

ML was applied using support vector machine (SVM) with two separate kernels, including polynomial and linear, and random forest (RF). RF was selected due to its effectiveness in classifying contractile profiles, while SVM was selected as the commonly used ML approach.

GraphPad Prism software (version 8.0.2) was used to perform statistical analysis. A two-way ANOVA was performed after the normal distribution of the data was checked using the Kolmogorov–Smirnov test, correlated with the Tukey test for several comparisons.

Research findings

The results demonstrated that all software investigated in this study produced comparable estimations of cardiac mechanical parameters. For instance, the beat duration measurements in cardioids at 0.5 Hz were 937.11 ms for ViKiE, 1043.59 ms for CONTRACTIONWAVE, and 1053.58 ms for MUSCLEMOTION.

ViKiE displayed a higher sensitivity in exposed samples due to its localized kinematic analysis, while CONTRACTIONWAVE and MUSCLEMOTION provided temporal correlation, integrating global assessment with time-efficient analysis. SVM with both polynomial and linear kernels and RF showed good performance on both MUSCLEMOTION and CONTRACTIONWAVE training datasets. ViKiE underperformed with SVM but demonstrated good performance with RF.

All classification models performed better with MUSCLEMOTION compared to ViKiE and CONTRACTIONWAVE. SVM with polynomial kernel underperformed, specifically when trained using ViKiEand CONTRACTIONWAVE data. RF demonstrated the best performance in both adult single ventricular cardiomyocytes and cardioids at higher frequencies, specifically one Hz for cardioids and two Hz for single cells. Overall, ML better classified MUSCLEMOTION/realized greater accuracy of more than 83% when trained with the MUSCLEMOTION dataset compared to the other two software programs.

In conclusion, the findings of this study offered crucial insights for the precise selection and integration of software tools into the kinematic analysis pipeline, tailored to the experimental protocol.