Deep Convolutional Networks for Automated Volumetric Cardiovascular Image Segmentation: From a Design Perspective
Ayman El-Baz, Jasjit S. Suri in Cardiovascular Imaging and Image Analysis, 2018
Previous studies [43] have shown that small convolutional kernels are more efficient in network design. The effective receptive field size of stacked small kernels is equivalent to that of one large kernel (the effective receptive field of three kernels is the same as one kernel), while giving lower computation cost. Therefore, we adopt small convolution kernels with size of in convolutional layers. Each convolutional layer is followed by a rectified linear unit (ReLU) as the activation function. Note that we also employ batch normalization layer (BN) before each ReLU layer to accelerate the training process. At the end of the network, we add a convolutional layer as a main classifier to generate the segmentation results and further get the segmentation probability map after passing the softmax layer.
Automatic Mask Detection and Social Distance Alerting Based on a Deep-Learning Computer Vision Algorithm
S. Prabha, P. Karthikeyan, K. Kamalanand, N. Selvaganesan in Computational Modelling and Imaging for SARS-CoV-2 and COVID-19, 2021
The architecture of the Convolutional Neural Network model consists of many convolution layers, pooling layers and fully-connected layers, as shown in Figure 5.1 (Huang et al., 2018). Filtering and smoothing operations are carried out in these layers of CNN. AlexNet is the basis of modern CNN, which is comprised of eight convolution layers, three pooling layers, and three fully connected layers, with Relu as the activation function. A plethora of improvisations, inspired by AlexNet, came into existence: VGGNet; GoogleNet; and ResNet. These accumulated convolution layers, pooling layers and fully connected layers (Szegedy et al., 2016). The purpose of generic object detection is to locate and identify current objects in any single image and to labeling them with rectangular bounding boxes to signify the confidences of existence (Uijlings et al., 2013).
Deep Learning Architecture Design for Multi-Organ Segmentation
Jinzhong Yang, Gregory C. Sharp, Mark J. Gooding in Auto-Segmentation for Radiation Oncology, 2021
As discussed previously, the CNN methods input the downsized input image or patch into the convolutional layers and fully connected layers, and subsequent output-predicted label. Shelhamer et al. first proposed a CNN whose last fully connected layer is replaced by a convolutional layer. Since all layers in this CNN are convolutional layers, the new network is named as a fully convolutional network (FCN). Due to the major improvement of deconvolution kernels used to up-sample the feature map, an FCN allows the model to have a dense voxel-wise prediction from the full-size whole volume instead of a patch-wise classification as in a traditional CNN [70]. This segmentation is also called “end-to-end segmentation”. By using an FCN, the segmentation of the whole image can be achieved in just one forward pass. To achieve better localization performance, high-resolution activation maps are combined with up-sampled outputs and then passed to the convolution layers to assemble more accurate output.
A review on patient-specific facial and cranial implant design using Artificial Intelligence (AI) techniques
Published in Expert Review of Medical Devices, 2021
Afaque Rafique Memon, Jianning Li, Jan Egger, Xiaojun Chen
Fuessinger et al. used an SSM method to reconstruct an artificial bone defect on the right temporal bone. Statistical Shape Models are geometric models that explain the collection of semantically similar objects in a compressed form. SSM represent an average shape of various 3D objects as well as their variation in shape. The actual target surface bone was very near but does not exactly the same because the shape variability of the SSM was not available [42]. However, a Convolutional Neural Network (CNN) is a class of neural networks that is well known for processing data that has a grid like topology, such as an image. A digital image is a binary representation of visual data which contains pixels arranged in a grid-like fashion that contains pixel values to denote how bright and what color each pixel should be.
Deep learning for assessing liver fibrosis based on acoustic nonlinearity maps: an in vivo study of rabbits
Published in Computer Assisted Surgery, 2022
Jinzhen Song, Hao Yin, Jianbo Huang, Zhenru Wu, Chenchen Wei, Tingting Qiu, Yan Luo
CNN is a deep feed-forward neural network model for processing data with mesh-like features. The sparse connection with the parameter weight sharing methods is integrated by CNN, which results in a significant reduction in the number of training parameters and effective avoidance of algorithm over-fitting. Meanwhile, the back-propagation algorithm is used to update the model parameters. A typical CNN structure consists of feature extraction network and classification network. Convolutional layers and pooling layers are the main components in feature extraction network. Convolutional layer contains a bunch of learnable filters (kernels) to convolve input data to generate feature maps as the input to the next layer. Once the linear transformations are completed, non-linear activation functions are applied to the output of filters. Behind the convolutional layer is the pooling layer. In pooling layer, the dimension and quantity of trainable parameters are diminished by pooling kernels. The process of pooling is aimed to control over-fitting and increase robustness to minor variations in the input data. In classification network, fully connected layer applies a linear transformation to extract and connect features. The softmax classifier is located behind fully connected layers to ensure that the output vector is normalized to sum to 1, reflecting an estimate of the posterior class distribution [25].
Developing neural network model for predicting cardiac and cardiovascular health using bioelectrical signal processing
Published in Computer Methods in Biomechanics and Biomedical Engineering, 2022
Sergey Filist, Riad Taha Al-kasasbeh, Olga Shatalova, Altyn Aikeyeva, Nikolay Korenevskiy, Ashraf Shaqadan, Andrey Trifonov, Maksim Ilyash
CNN eliminates the main disadvantage of FNN—the formation of a space of informative features. In CNN, it is generated automatically, but here, too, a number of problems arise related to the empirical selection of CNN meta-parameters. Compared to traditional machine learning methods that separate feature extraction and classification, CNN can automatically extract powerful high-level features and implement classification based on them, which means that CNN combines these two stages of image classification. CNN usually consists of a convolutional layer, an unification layer and a fully connected layer (FC) (Al-Kasasbeh et al., 2019a). Among them, the convolutional layer automatically extracts task-related functions, and the generalizing layer performs the task of descriptor reduction (downsampling). The fully connected layer calculates the confidences for each class. In Wang et al. (2020), a ten-layer CNN is presented, designed to diagnose patients with alcoholism from images of the brain obtained with an MRI scanner. CNN uses a new model of the activation function and additionally introduces a batch normalization block and a dropout block. The results show that the method achieved a sensitivity of 97.73 ± 1.04%, a specificity of 97.69 ± 0.87% and an accuracy of 97.71 ± 0.68%.
Related Knowledge Centers
- Cerebral Cortex
- Time Series
- Visual Cortex
- Medical Image Computing
- Brain–Computer Interface
- Overfitting
- Visual Field
- Receptive Field
- Feature
- Simple Cell