ScaledDenseNet: An Efficient Deep Learning Architecture for Skin Lesion Identification

CNN has been at the forefront in the field of object recognition and image classification since the creation of alexnet. The feature matrix is downsampled via max pooling after numerous convolutional layers are used to extract the features. Additionally, it makes use of overlapping pooling for improved feature extraction during downsampling and ReLU nonlinearity for effective training [1]. There have been numerous attempts to improve convolutional neural net accuracy by deepening it [2-6]. Increasing network depth was a priority for the VGG network. There was a total of 6 network architectures suggested. The accuracy findings were clearly improved by scaling the network depth [2].

Multiple networks have used the concept of inception in image classification. GoogleLeNet uses inception to achieve sparse connectivity. Inception also helps in reducing training complexity as only a part of the network is trained and reused. GoogleLeNet also acknowledges that increasing network depth is the most straightforward way to increase network accuracy. Inception net uses 1x1 convolutions as a way of dimensionality reduction and reducing training complexity. It uses inception modules, which are layers of parallel convolutional layers [3].

Simply increasing the network depth leads to training inefficiencies as well as a major roadblock known as the vanishing gradient problem. Due to the nature of the gradient descent algorithm, increasing network depth will lead to smaller gradients being passed down to the previous layers. As the gradient approaches zero, the layer will completely stop training. To overcome this problem, various residual networks have been proposed [4, 5].

ResNet uses identity mapping to overcome vanishing gradients. This allows the network to reach very deep. Identity mappings or residual mappings are also called skip connections or shortcut connections. In ResNet, the shortcut connections are just the identity mappings between the input and output of a layer. Going deeper allowed for an increase in the accuracy of the classification task [4].

Similar to ResNet, DenseNet also uses the concept of reusing features. DenseNet consists of multiple dense blocks and transition layers in between these dense blocks. In a dense block, the output of each layer is connected to the input of every other successive layer in the block. In order to avoid dimensional discrepancy, downsampling is only done after a dense block. This allows the network to go deeper without vanishing gradient. Feature reuse certainly helps in achieving better accuracy compared to residual mapping [5]. Increasing the network depth, width and resolution helps in increasing accuracy of the classification task. However, there hasn't been a clear method to bind these 3 factors together. Google’s EfficientNet introduces the concept of compound scaling to increase network depth, width, and resolution simultaneously. Increasing one of these parameters quickly leads to the saturation of results where the accuracy does not increase anymore. Compound scaling overcomes this by simultaneously being more efficient [6]. There have been attempts at using CNNs to predict skin disease, particularly, skin cancer, using histologic images. Höhn et al. [7, 8] used patient metadata such as gender, lesion location, age, etc in combination with images to predict the chances of a patient having a particular type of disease. For unbalanced types of data, metadata increased the accuracy. In both cases, two types of methods to combine metadata were used. One method was to concatenate the metadata vector to the feature vector of the input image before the classification layer. Another was to scale the feature vector using a metadata vector. The latter turned out to be better than the former method. Höhn et al. [7] used ResNext50 as the feature extraction CNN. Li et al. [8] tested the metadata combination method on multiple state-of-the-art CNNs. Kaur et al. [9] proposed a new CNN architecture that uses multiple parallel blocks of convolutional layers. The network has 11 primary blocks. Each block uses normalisation to improve accuracy and leaky ReLU to overcome vanishing gradient problems. The dataset was obtained from ISIC 2016 and ISIC 2017 challenges. The network classifies melanocytic lesions and benign lesions. Various data augmentation techniques such as rotation, scaling, etc. were used to overcome the imbalanced training dataset problem. Raza et al. [10] used ensemble technique for the classification of skin cancer. A stack-based approach for ensembling is proposed over the traditional approaches. The ensembled networks are trained using transfer learning before ensembling. VGG16, Xception, InceptionResnetV2, DenseNet121, DenseNet169, and DenseNet210 are fine-tuned for melanoma classification. Nithya Anoo et al. [11] proposed a new CNN architecture which has alternating convolutional and max pool layers. ALEnezi [12] used a pretrained AlexNet for feature extraction and SVM for classification of 4 types of skin diseases. Kumar and Kumar [13] used a machine learning approach to the feature extraction problem. The input images are first segmented using the active contour method and the features are extracted using mathematical equations to identify brightness, colour, lesion size, etc. The classification was done using an ANN. Mahbod et al. [14] used various pre-processing techniques such as colour standardisation and normalisation for better feature extraction. Various state-of-the art neural networks were fine-tuned for the melanoma classification task. An ensemble of these networks was used to classify the skin cancer images. Various machine learning techniques have been used for classification tasks as well. Pathan et al. [15] used color features of the dermatoscopic images as features for the classification problem. A total 15 features were extracted from each image. Based on these 15 features, the images were classified using a decision tree algorithm.

Methods explained as above are either using an ensemble of networks or baseline network architecture for the purpose of classification of dermatoscopic images. However, ensemble methods are multiple times more computationally intensive than plain baseline networks. The accuracy gains from ensemble methods are marginal. Most baseline networks are also limited by the fact that they only scale a single dimension, i.e., network depth. This makes the result saturate quickly when increasing the network depth. The proposed architecture makes use of efficient compound scaling and DenseNets property of feature reuse for efficient but an effective classification model.

4.1 System design

Datasets are divided as training and testing categories. Then, the training datasets were pre-processed by resizing the images to appropriate dimensions. Pre-processing training data is passed to proposed architecture. The proposed architecture trained on training data and classified each image to proper classes according to accuracy. After completing the model's training process, the proposed model was tested by passing images from the test data to check whether each image shows the correct class. The procedures for the Model ScaledDenseNet are presented in Figure 1.

figure1.png

Figure 1. Model ScaledDenseNet architecture procedure diagram

4.2 Compound scaling

Compound scaling, introduced through the EfficientNet concept, offers several benefits for skin disease classification. It balances the scaling of network depth, width, and input resolution, resulting in improved classification performance. Scaling the network depth enables the capture of complex patterns and prevents overfitting. Increasing network width enhances the model's ability to learn and model variations in the data. Scaling the input resolution allows for capturing finer-grained features and textures. By simultaneously scaling these dimensions using a compound scaling coefficient, the classification process is optimized, leading to a more accurate and reliable skin disease classification.

depth: $d=\alpha^\phi$    (1)

width: $w=\beta^\phi$    (2)

resolution: $r=\gamma^\phi$    (3)

The values of α, β, and γ are hyperparameters and ϕ is a user-defined scaling factor.

The increase in the number of operations due to scaling will be proportional to α×β²×γ² [6].

The value of hyperparameters α, β, and γ is limited such that α×β²×γ² is less than or equal to 2.

4.3 Dense blocks

The dense block is a concept introduced in DenseNet architecture. Dense blocks are used to overcome the vanishing gradient problem. Successive convolution layers in dense blocks take input from each of the previous layers. It establishes direct connections between layers, preserving information and gradients, leading to efficient gradient flow and improved learning. The dense connections enable extensive feature sharing, allowing the network to capture both low-level and high-level features. The information is preserved due to this. Dense blocks also use compression factor ө for parameter reduction. The value of compression factor ө lies between 0 and 1. The compression factor reduces the number of output layers on successive layers. Dense blocks also have a bottleneck layer before the actual convolution layer to reduce the number of features. Table 2 presents a detailed view of the architecture of ScaleDenseNet when ϕ=3.

Table 2. Architecture of ScaledDenseNet (ϕ=3)

Layers	Output Size	Layer Size	Number of Kernels
Input	160x160	-	-
Convolution	80x80	7x7 conv	67
Pooling	40x40	3x3 max pool	-
Scaled Dense Block X 3	40x40	1x1 conv	134
		3x3 conv	33
Transition Layer I	20x20	-	-
Scaled Dense Block X 6	20x20	1x1 conv	134
		3x3 conv	33
Transition Layer II	10x10	-	-
Scaled Dense Block X 13	10x10	1x1 conv	134
		3x3 conv	33
Transition Layer III	5x5	-	303
Scaled Dense Block X 9	5x5	1x1 conv	134
		3x3 conv	33
Global Average Pool	1x1	5x5 Global Avg Pool	-
Classification		7D fully connected

Table 3. Transition layers of ScaledDenseNet (ϕ=3)

Layers	Layer Size	Number of Kernals
Transition Layer I	1x1 conv	87
	2x2 Avg Pool	-
Transition Layer II	1x1 conv	149
	2x2 Avg Pool	-
Transition Layer III	1x1 conv	303
	2x2 Avg Pool	-

figure2.png

Figure 2. ScaledDenseNet block diagram

4.4 Transition layers

This is also a concept introduced in DenseNet. Since the feature map size should not change across the dense block, the downsampling is done in the transition layer by using average pooling layers. Similar to dense blocks, the transition layers also have a bottleneck convolution layer before actual downsampling. Table 3 displays the comprehensive design of ScaleDenseNet's transition layer architecture.

The proposed architecture combines the compound scaling of EfficientNet with the feature reusability of DenseNet. It is observed that balancing the depth, width, and resolution of the network is essential to increasing accuracy [6]. EfficientNet provides a way to tie these three factors together. DenseNet uses feature reuse as a way to overcome the vanishing gradient problem. Figure 2 displays a block diagram of the ScaledDenseNet model. The diagram presents an overview of the model's building blocks, including the Convolutional layer, ScaledDense Blocks, Transition Layers, and Pooling layer. By visualizing the components of the model and their connections, Figure 2 provides a valuable resource for understanding the architecture and design of the model.

This allows DenseNet to go deeper with increased accuracy [5]. These features of EfficientNet and DenseNet are complementary to each other. Hence, we have chosen to apply the compound scaling technique to DenseNet. We will call this architecture ScaledDenseNet. Similar to EfficientNet, our proposed architecture used d, w, r as the scaling factors to scale depth, width, and resolution respectively [6].

where, $d=\alpha^\phi$ , $w=\beta^\phi$  and $r=\gamma^\phi$

The values of α, β, and γ are hyperparameters and ϕ is a user defined compound scaling coefficient. The values of α, β, and γ are found by setting ϕ=1 and performing a grid search. The values of hyperparameters were restricted such that α×β²×γ²≤2. Since the increase in the number of operations will be proportional to α×β²×γ² [6].

For the base network, we used the DenseNet-121 [5] architecture. Using d, only the number of layers in each dense block is scaled. Using w, the number of filters in every layer is increased, while r is used to scale the input resolution only.