Why HarmonyCVI?HarmonyCVI uses Tech Vedika’s Vision Analytics platform to analyze a MRI scan from all points of the cardiac life cycle in under 2 minutes, thereby saving 28 minutes of the clinician’s time, every time. (The current manual process takes over 30 minutes to analyze the scan for 2 points of the cardiac life cycle).
Functions of HarmonyCVI
- Volumetric Analysis
- Q-Flow Analysis
How HarmonyCVI works?
Our AI-powered Vision Analytics platform for analyzing cardiovascular images has outperformed other similar state-of-the-art methods in terms of accuracy, robustness, and computational time.
100% Cloud Based
- 24×7 cloud-based global access, eliminating the need to manage software and hardware. Significant cost savings and greater flexibility
- Scalability-on-demand through addition of more Graphical Processing Units (GPUs) ensuring 100% uptime
Trained on 1000’s of images
HarmonyCVI algorithms are pre-trained; they are productive from Day 1.
Simple User Interface
- Simple and user friendly interface
- Quick to learn, deploy and use
- Deliver powerful insights
- Allows for collaboration with other clinicians
HarmonyCVI uses the most advanced web rendering technology for 2D/3D image visualization, appropriate annotation and highlighting of critical aspects. This technology employs state-of-the-art graphics to deliver a comprehensive view.
The reports help clinicians to quickly assess the key parameters for quick diagnosis of cardio-vascular conditions.
- Full ventricular and atrial assessment
- Assessment of atrioventricular junction
- Contour detection
- Thresholding tool
- Comparison of baseline with follow-up scans
- Drag and drop images
- Multiple export formats including DICOM encapsulated PDF
- Detailed measurements of key statistics
17 Segment Model*
- Creates a distribution of 35%, 35%, and 30% for the basal, mid-cavity and apical thirds of the heart
- Cardiac segmentations and their assignment to coronary arterial territories
HarmonyCVI uses a Fully Convolutional Neural Network (CNN). It takes the MRI images as input, learns image features from fine to coarse scales through a series of convolutions, concatenates multi-scale features, and finally performs a pixel-wise image segmentation for cardiovascular analysis.
The dataset consists of short-axis and long-axis cine MRI images. Manual image annotation is undertaken by a team of clinicians.
For short-axis images, the LV endocardial and epicardial borders and the RV endocardial borders are manually traced at ED and ES time frames.
For long-axis 2-chamber view (2Ch) images, the left atrium (LA) endocardial border are traced. For long-axis 4-chamber view (4Ch) images, the LA and the right atrium (RA) endocardial borders are also traced.
The Cardiovascular DICOM images are converted into Neuroimaging Informatics Technology Initiative (NifTI) format. The manual annotations from the CVI software are exported as XML files as well as in NIfTI format.
Automated Image Analysis
For automated CMR image analysis, HarmonyCVI utilizes a fully convolutional network architecture, which is a type of neural network that can predict a pixel-wise image segmentation by applying several convolutional filters onto an input image.
The fully convolutional network learns image features from fine to coarse scales using convolutions and combines multi-scale features for predicting the label class at each pixel.
The network consists of several convolutional layers for extracting image features. Each convolution uses a 3X3 kernel and it is followed by batch normalization and rectified linear unit (RELU). After every two or three convolutions, the feature map is downsampled by a factor of 2 so as to learn features at a more global scale.
Feature maps learned at different scales are upsampled to the original resolution using transposed convolutions and the multi-scale feature maps are then concatenated. Finally, three convolutional layers of kernel size 1X1, followed by a softmax function, are used to predict a probabilistic label map. The segmentation is determined at each pixel by the label class with the highest softmax probability. The mean cross-entropy between the probabilistic label map and the manually annotated label map is used as the loss function.
Network Training and Testing
Networks are trained for segmenting short-axis images, long-axis 2Ch images and 4Ch images. Data augmentation is performed on-the-fly, which includes random translation, rotation, scaling and intensity variation to each mini-batch of images before feeding them to the network.
Evaluation of the Method
HarmonyCVI’s performance was evaluated and the results were comparable/better than competing models using commonly used metrics for segmentation accuracy assessment, such as Dice metric, mean contour distance and Hausdorff distance, and clinical measures derived from segmentation, including ventricular volume and mass.