Table 4.
Different DCNN models trained in this study and model descriptions.
| Model | Description | Number of layers/trainable parameters |
|---|---|---|
| VGG16 | A 16-layer architecture consisting of convolution layers, Max-pooling layers, and 3 fully connected layers at the end. It has a deep network but end-to-end small 3 × 3 Convolutional filters |
16 layers 138.4M parameters |
| DenseNet201 | A CNN architecture consisting of Densely connected blocks, where each layer input comes from previous layer output feature maps. It has two block types, Dense blocks including batch normalization, ReLU activation and 3 × 3 convolution layers, a Transition layer consisting of Batch normalization, 1 × 1 convolution and Average pooling layers. Transition blocks are placed after each dense blocks |
402 layers 20.2M parameters |
| MobileNet | An architecture that utilizes depth-wise separable convolutions and thus reducing the number of parameters. These are made of two operations: depthwise convolution for filtering, and point-wise convolution for combining the outputs of depth-wise convolutions with 1 × 1 convolution |
55 layers 4.3M parameters |
| ResNet152 | The main feature of ResNet architecture is the existence of residual blocks that utilize shortcuts to skip some layers. Each residual block consists of two Conv-layers, with batch normalization and ReLU activation, using 3 × 3 filters with stride 1. Resnet is famous for solving the Vanishing Gradient problem |
307 layers 60.4M parameters |
| InceptionV3 | A CNN model that is made of symmetric and asymmetric building blocks that consist of Convolutions, AVG-pooling, Max-pooling, dropouts, and fully connected layers. The convolutions are factorized that results in a reduced number of learnable parameters |
189 layers 23.9M parameters |
| NASNetLarge | Stands for Neural Search Architecture network and works best on small datasets. In simple terms, it automates the network architecture engineering, and identifies and evaluates the performance of possible architecture designs without training. Furthermore, it utilizes a regularization technique called ScheduledDropPath |
533 layers 88.9M parameters |
| CheXNet | It is a 121 layer Convolutional neural network that inputs a chest X-ray image and outputs the probability of a pathology |
121 layers 6.9M parameters |