Fig 1.
The steps of proposed method of RWMA detection using echocardiography images for RWMA detection, consisting with cardiac structure segmentation using a trained U-Net on both A4C and A2C echo frames, resulting in segmented echo frames; motion features engineering with optical flow frames in both the A4C and A2C views, with backbone Temporal ConvNets to analysis the X and Y components of the motion; RWMA detection, using extracted features are fed into a suite of ML classifiers—including KNN, DT, Random Forest, SVM, and MLP—to achieve binary classification for the presence of RWMA or non-RWMA conditions.
Fig 2.
Selection of the study population from the HMC-QU dataset for temporal ConvNet backbones finetune and RWMA detection classifiers training and validation.
Table 1.
The number of subjects with respect to their corresponding ground-truth labels from A4C and A2C views.
Fig 3.
ROI of cardiac structure in echocardiograms generated by U-Net.
Fig 4.
(a) A2C and (b) A4C displacements curves recordings of a patient consist of two cardiac cycles from A2C and A4C echocardiograms.
Fig 5.
The motion feature extraction part for the Temporal ConvNet backbone employed for echocardiograms.
Each of the K optical flow format echocardiogram snippets is classified using one Temporal ConvNet instance. The horizontal(X) and vertical(Y) Optical flow fields were extracted 5 frame each time. The first convolution layer is adjusted to 224×224×10 to accept. The snippet-level echocardiograms c (s;) are aggregated in the consensus layer, yielding the overall RWMA detection. Snippets are stacks of several optical flow fields calculated between consecutive echocardiogram frames. Temporal ConvNets backbone on all snippets share same parameters.
Fig 6.
Overall confusion matrices of all classifiers used in our study for RWMA detection performance results (multi-vies) with the optimized parameters by grid search computed over the test sets of 5−fold in HMC-QU dataset.
Table 2.
The Performance of pre-trained U-Net model on the HMC-QU dataset.
Fig 7.
Mean AUC for RWMA detection performance results (multi-vies) computed over the test sets of 5−fold in HMC-QU dataset.
Table 3.
The average RWMA detection performance results (multi-views) with the optimized parameters by grid search.
Table 4.
Overall RWMA detection performance results (multi-views) computed over the test sets of each 5−fold in HMC-QU dataset.
Table 5.
Average RWMA detection performance results (A4C view) computed over the test sets of each 5−fold in HMC-QU.
Table 6.
Average RWMA detection performance results (A2C view) computed over the test sets of each 5−fold in HMC-QU.
Table 7.
Comparison of related works by using HMC-QU dataset.
Fig 8.
A comparative generated dense optical-flow frames of the original echocardiogram versus the U-Net processed echocardiogram.
Representation includes the sample frame of input video, the generated optical flow frame, which are visually enhanced with colour arrows to indicate the direction and magnitude of motion, for both (a) Apical 2 View and (b) Apical 4 View.