Empirical Investigations to Skin Lesion Detection Using DenseNet Convolutional Neural Network

• Ghalejoogh et al. [10] presented an automated skin lesion detection system. In order to remove hair from the lesion images pre-processing was employed. Then the lesion image segmentation is done using Otsu thresholding. Feature extraction is done based on the colour, shape and texture. Feature selection is done using Wrapper methods. Hierarchical structure-based stacking is used to classify the skin lesions
• Al Masani et al. [11] performed the evaluation of skin lesion CAD. He proposed that more images. The accuracy of classification also improved when the model is fed with the segmented lesion images. It proved ResNet-50 provided a better accuracy when compared to other CNN models.
• Xie et al. [12] presented a model to classify the skin lesion as malignant or benign. The model is deployed to work on the incomplete lesion images present in the dataset. Dimensionality reduction is used to eliminate the noisy features. Dimensionality reduction used is PCA. An ensemble model of BP network and fuzzy network is used for classification.
• Hekler et al. [13] presented a model combining the human intelligence and artificial intelligence. A CNN model is trained with 11,000 images and 117 doctors were used to classify the skin lesions. This diagnosis is combined into a classifier. The computation time is higher for the model developed.
• Sae-Lim et al. [14] presented an MobileNet CNN model. MobileNet is used for classification of the skin lesions. The efficiency of the classifier is improved by data up sampling and data augmentation. The model is tested on the HAM10000 dataset. The model gave better accuracy and precision and f-score compared to existing methods.
• Kim et al. [15] presented an approach for automatic detection of the salient regions. The saliency map is constructed using a linear combination of colours. The trimap is constructed overcomes the limitation of the saliency map. The model performed better on three datasets.
• Chen et al. [16] presented a method called ESSL which is derived from ELM classifier to classify the multiple classes of skin lesion. ESSL performed better than the SVM. This model can handle the memory issue. The model can classify the skin lesions in large datasets.
• Tang et al. [17] presented a Global part CNN model. The model treats the global and local information equally. The G-CNN model is used to extract the global information of dermoscopy images. The P-CNN model is used to extract the local information of the lesion images.
• Afza et al. [18] presented a model for skin lesion detection using statistical normal distribution and optimal feature selection. Statistical normal distribution is used for segmentation of lesion images. Histogram, colour are the extracted features. The selected features are fed into a CNN model. Best accuracy is obtained for cubic function.
• The existing models tested their accuracy with only one dataset and the models may not work properly for imbalanced datasets. The models take more time for computation. A model is proposed to perform accurately on all the datasets and must work on imbalanced datasets.

4.1 Proposed methodology

The Entire process of the proposed work can be seen in Figure 9. In this work, input data is taken from the “HAM10000” dataset from Kaggle. In the next step, input images are resized using the computer vision library, resizing of images must be done to focus more on the lesion area, splitting of the dataset into training dataset and testing dataset is done. In the next step, 80% of the dataset is used for training and 20% is used for testing. Here we proposed DenseNet Model with five convolutional layers is trained up to 100 epochs by training dataset. The trained DenseNet model is tested on the testing dataset and the accuracy is measured and evaluated. Our experimental investigations emphasize that the detection of skin lesion of input data image.

9.png

Figure 9. Block diagram for proposed methodology

4.2 Modules

Importing the required libraries:

import numpy as np

import pandas as pd

import matplotlib.pyplot as plt

from tensorflow. Keras. Preprocessing. Image import ImageDataGenerator

• NumPy:

NumPy is a predefined library in python. Numerous mathematical operations can be carried out on arrays with NumPy.

• Pandas:

Pandas is an open-source library designed primarily for using relational or labelled data quickly and easily.

• Matplotlib:

For 2D displays of arrays, the Python visualization tool Matplotlib is fantastic.

• Tensorflow:

TensorFlow, an Open-Source deep learning and machine learning library, is useful for voice search, text-based applications, image identification, and many other things.

• ImageDataGenerator:

ImageDataGenerator is used to transform the input image like rotating, resizing etc.

Reading the input data:

With the help of pandas, we can read the HAM10000 dataset which is required in our project. HAM10000 (Human Against Machine) is an image dataset which is published by Vienna Medical University. This dataset is adopted for skin lesion classification as seven different types of classes. This dataset contains of about 10,015 skin lesion image samples which are collected from different age groups and different parts of the patients.

Resizing the images in the dataset:

The most crucial step is resizing the photographs. The dataset's original image will be 500 x 500 pixels in size. The lesion area, the healthy area, and the skin's hair can all be seen in this photograph. The photograph needs to be enlarged in order to highlight the lesion region more. The original image has been resized to 100 × 100 in size. Resizing places additional emphasis on the lesion region. The accuracy of the classification is impacted by this step. It is necessary to resize and store in an array all 10,015 photos. With the use of computer vision, the image is resized. The fundamental segmentation technique we use is image resizing. By this step, we focus more on the lesion area by leaving the unwanted skin parts. The output of the entire resizing image can be seen in Figure 10.

10.png

Figure 10. Output of resizing the image

Splitting the dataset:

The splitting of dataset as training dataset and testing dataset is done in this step. The training dataset will be used for training the DenseNet CNN model. The trained DenseNet model will be tested on the testing dataset. In order to avoid underfitting and overfitting, 80% of the dataset will be used as training dataset and 20% will be used as testing dataset.

The size of the total dataset is 26,115 images of dimension 100 x 100.

The size of our training set is 20,892 images.

The size of our testing dataset is 5,223 images.

Training DenseNet CNN model:

The architecture of DenseNet is mentioned in the below Figure 11. The two types of layers convolution and pooling are alternative. The number of filters increase as the network moves from left to right. The last stage of the network contains one or more fully connected layers which classifies the output. The architecture of DenseNet consists mainly four types of layers convolutional layer, dense block, pooling layer and fully connected layer.

11.png

Figure 11. DenseNet CNN architecture

• Dense Block: Dense Blocks links all layers directly with matching feature-map sizes. In order to maintain the feed-forward nature, each layer receives extra inputs from all levels that came before it and transmits its own feature-maps to all layers that came after it.
• Covolution: The use of convolution is to identify the suitable attributes from the image that is taken as the input.
• Pooling: Pooling is also known as down sampling. It is used in the reduction of the unwanted features that are extracted.
• Fully connected: The classification is done by the fully connected layer. The outputs of convolution and pooling passes to fully connected layer. These outputs are used to classify the input image.

• ReLu: ReLu means “Rectified Linear Unit”. “ReLu” activation function is used to make the negative features extracted by the model as zero. If the features extracted are positive then those features will be sent to the next layers. If the features are negative then they will made as zero.
• Batch Normalization: Every layer of the network can learn more independently thanks to a layer called batch normalization. It is used to normalise the output of the preceding layers. The normalising activations scale the input layer. Usage of batch normalization makes learning more effective.
• Dropout: The regularization method used to stop overfitting in the model is called dropouts. A certain percentage of the network's neurons are switched at random with the addition of dropouts. The incoming and outgoing connections to the neurons are also turned off when they are turned off. To help the model learn more effectively, this is done.
• SoftMax: SoftMax is used to label the regions present in the image. The lesion area is labelled as ‘1’ and the healthy area is labelled as ‘0’. SoftMax layer is the output layer. It gives the classified output image.

The DenseNet model is trained on the training dataset. At first input image of dimensions 100 x 100 x 3 is given as input to covolution layer. The number of filters is 96 in first convolution layer. The output of the first convolution layer is sent to max pooling layer for down sampling. The second convolution layer takes the output of max pool layer as input. It contains 256 filters. These operations are performed until the image is flatten out. The dropout factor is 0.5. The activation function used in hidden layers is ‘ReLu’ and activation function used in output layer is ’SoftMax’.

Testing the model:

The trained model is tested on the testing dataset which contains 5,233 images. The model accurately predicted 4,244 images.

Measuring the accuracy:

There are seven classes of lesions in the HAM10000 dataset. They are:

• Melanocytic nevi
• Melanoma
• Benign-Keratoses
• Basal cell carcinoma
• Actinic keratoses
• Vascular lesions
• Dermatofibroma

The accuracy of each class is measured as per the predicted output of the lesion. If the predicted type matches with the type of the lesion present in the dataset the accuracy is counted.

Skin Lesion Detection:

The output of the model is the prediction of the type of the lesion. The result is displayed in the Figure 12.

12.png

Figure 12. Skin lesion detection