A Distributed Densely Connected Convolutional Network Approach for Enhanced Recognition of Health-Related Topics: A Societal Analysis Case Study

2.1 Datasets used previously

The systems are intended to be learned from one or more Data Sets. Based on existing publications in the literature, our research intends to demonstrate the expanding tendency of such self-diagnosing automated systems, segmenting or detecting certain Skin Lesions (especially melanoma) available in Table 1 and Table 2.

Clinicians examine the lesions visually using the ABCDE criteria [13], followed by histological examination. Due to the comprehensive understanding of biological models, many algorithms [14] relied mainly on craft feature sets that had low aggregation for dermoscopic images.

Table 1. Frequently used datasets of skin lesions

2.2 Melanoma detection, segmentation, and classification using neural networks

Deep Convolutional Neural NetworksCNNs, NNs and TL for NNs are used majorly. According to the most recent research on Skin Disease prediction, segmentation, and classification publications in the literature. As a result, NNs are used in the vast majority detection of Melanoma, segmentation, and classification systems.

Table 2. Neural networks used majorly

2.2.1 GoogLeNet/Inception

GoogLeNet, commonly known as Inception v1. GoogLeNet's simplified architecture has twenty-two layers using a Softmax layer of Thousand neurons. Another essential point to note is that ReLU is used as an activation function by all convolutions within the design.

2.2.2 Xception network

Another CNN used to diagnose skin lesions is Xception, which is designed to Replace standard Inception modules with depth-separable curves to boost performance. The superior to Inception v3 Xception architecture consists of 36 convolutional layers grouped in 14 modules, all of which are surrounded by residual connections [26].

2.2.3 DenseNet

DenseNet is a family of CNNs commonly used in the diagnosis of skin lesions, all of which were published in 2020, are examples of publications that use DenseNet (especially DenseNet-201). Thanks to its efficiency and higher ACC, DenseNet is a recent trend in recently published articles. The reason for this is that in the early work [27] the authors included tightly bound layers, which changed the usual design of CNN. DenseNet networks published in various studies include DenseNet-121, DenseNet-169, DenseNet-201 and DenseNet-264.

2.2.4 NASNet

“A convolutional neural network named NASNet-Large was developed and trained using more than a million photos from the ImageNet collection” [1]. The network has 1000 item categories it may use to categorise photos. NASNet (Neural Architecture Search Network) models using weights which have already been trained on ImageNet. For the NASNetLarge model, the default input image size is 331×331; for the NASNetMobile model, it is 224×224.

In the year 2019, the main goal is to Pipeline architecture for SL segmentation, combining YOLO v3 and the GrabCut algorithm. This novelty can be attained by combining YOLO v3 and the GrabCut Algorithm for SL segmentation. Here the type of NN used is YOLOv3/ detection and segmentation. This was taken from the dataset PH2, ISIC 2017. In 2019, The goal is New FCNN architecture for SL segmentation-DermoNet because it describes that FCNN contains densely connected convolutional blocks and skip connections. Here type of NN used is FCNN-DermoNet/ segmentation from the dataset PH2, ISIC 2016, ISIC 2017.

In the year 2020, New deep CNN-based model for face skin disease classification using a triplet loss function is the novelty that can be described by Fine-tuning layers of ResNet152 and InceptionResNet-v2 by using ResNet152, Inception ResNetv2/classification function of NN. This dataset was taken from a hospital in Wuhan China.

Also in 2020, the novelty is Me detection using an optimized set of Gabor-based features and a fast MNN classifier. Here Gabor features combined with a fast (Multi-Level Neural Network) MNN and the type of NN used is MNN/classification. This dataset was taken from PH2.

In the year 2021, the novelty is to Design of a new DCNN model with multiple filter sizes-Classification of Skin Lesions Network (CSLNet). Here Fewer filters, parameters, and layers to improve SL classification performances. The NN function used is DCNN (CSLNet)/classification. This dataset was collected from ISIC 2017, ISCI 2018, ISIC 2019.In 2021, the goal is to Test different NN for recognition of pigmented SL by using ResNet50, DenseNet121, VGG16/classification. This dataset was collected from ISIC, HAM10000, PH2, BCN20000, SKUNK2.

In the year 2021, Combining the MobileNetV2 with the Spiking Neural Network (SNN) into a DCNN for the classification is the main goal where Three NNs connected into an intelligent decision support system for skin cancer detection. Here the type of NN used is Autoencoder, MobileNetv2, SNN/classification. This dataset was collected from ISIC [17].

To propose a fully automatic method for classifying several skin cancers by fine-tuning the deep learning models VGG16, ResNet50, and ResNet101. Prior to model creation, the training dataset should undergo data augmentation using traditional image transformation techniques and Generative Adversarial Networks (GANs) to prevent class imbalance issues that may lead to model overfitting [14]. With appropriate data augmentation, the proposed models attained an accuracy of 92% for VGG16, 92% for ResNet50, and 92.25% for ResNet101, respectively. The ensemble of these models increased the accuracy to 93.5%.

To stated that the limited size and shortage of diversity of accessible datasets of dermatoscopic images hampered intensive training of neural networks for automatic detection and classification of pigmented skin lesions, which had been addressed by the release of the HAM10000 [15].

To proposed method made use of transfer learning with a pre-trained AlexNet. The parameters of the original model were used as initial values, and the weights of the last three replaced layers were randomly initialised. The proposed method was a huge success because it accurately classified skin problems into seven classes. For accuracy, sensitivity, specificity, and precision, the attained percentages are 98.70%, 95.60%, 99.27%, and 95.06%, respectively [18].

To propose a unique category of fully convolutional network in this paper, along with new dense pooling layers, for segmentation of lesion regions in non-dermoscopic images. This proposed network, with exception of other established convolutional networks, is intended to generate dense feature maps [19]. This network results in highly accurate lesion segmentation. The produced dice score is 91.6%, which outclasses state-of-the-art algorithms in skin lesion segmentation using the Dermquest dataset.

To propose a systematic overview of recent advances in an area of growing involvement for cancer detection, with a concentrate on a contrasting perspective of cancer detection using artificial intelligence, particularly neural network-based systems. Such structures can be thought of as intelligent dermatologist support systems [20]. Theoretical and applied contributions to the major development trends of multiple neural network architecture based on decision fusion were investigated.

The study proof of dermoscopy's accuracy in diagnosing pigmented skin lesions, notably melanoma. The authors go through the numerous dermoscopy diagnostic criteria for melanoma as well as the value of dermoscopy training and experience [14]. In addition to highlighting the limits of dermoscopy in the identification of nonmelanocytic pigmented lesions, the study also discusses the possibility of computer-aided diagnosis (CAD) systems to increase diagnostic accuracy.

A skin cancer melanoma diagnosis machine learning project typically involves the following steps:

4.1 Data collection

The first step is to gather a dataset of skin lesion images, which includes both malignant and benign cases.

ISIC is an academic industry collaboration aimed at facilitating the use of skin image to help reduce melanoma death rate. ISIC achieves its goals by developing and promoting digital skin imaging standards and by collaborating with the dermatological and computer science communities to ensure diagnosis. ISIC is establishing suggested standards to improve the quality, privacy, and interoperability of digital skin scans ISIC is techniques for the computer science communities and dermatological, such as a massive and developing open source public access skin imaging archive.

4.2 Data pre-processing

The skin lesion images are pre-processed to enhance the contrast, remove artifacts, and standardize the size of the images.

4.2.1 Multi crop

To multi-crop, divide the input image into multiple sub-images, feed them into a network for classification, then average the results to improve accuracy.

• Five Crop divides the image into the four corners and a core crop.

• Ten Crop divides it all into 4 corners and the centre crop, as well as their inverted counterparts.

4.3 Feature extraction

Various image processing techniques are applied to extract features such as color, texture, and shape from the skin lesion images.

4.4 Model development

ML models are trained on the extracted features to classify the skin lesion images into malignant or benign classes. There are several ML algorithms that can be used for this purpose, including deep learning neural networks.

4.4.1 Algorithms

We will implement the DenseNet-161 version of the model

Bn-Relu-Conv Function

Algorithm:

Algorithm(bn_rl_conv(x,filters,kernel_size)

{

Int X;

      X=BatchNormalization(x)

      X=ReLU(x)

      X=Conv2D(filters=filters,

                          kernel_size=kernel_size,

                          padding=same)(x)

return X

Dense Block

Algorithm:

Algorithm(dense_block(tensor,k,reps):

{

Int x;

     for_in range(reps):

     x=bn_rl_conv(tensor,filters=4*k,kernel_size=1)

     x=bn_rl_conv(x,filters=k,kernel_size=3)

     tensor=Concatenate()([tensor,x])

}

Return tensor

4.4.2 Multi model

A multimodel of machine learning algorithms typically refers to the combination of multiple models, each utilizing a different algorithm, to create a more accurate and robust overall model. Algorithms like DenseNet and Xception are both popular deep learning architectures that can be used as base models in a multimodel approach.

In the case of algorithms like using DenseNet and Xception in a multimodel approach, one possible strategy would be to train each model independently on the same dataset, and then combine their predictions using an ensemble approach. Alternatively, the models could be combined into a single multimodal architecture, with each model feeding into a shared output layer. This approach is called a multi-input model and can be useful when the models have complementary strengths and weaknesses.

4.5 Model evaluation and validation

The performance of the ML model is evaluated using metrics such as accuracy, sensitivity, specificity, and area under the receiver operating characteristic curve (AUC-ROC). The model is also validated using a separate dataset to test its generalization capability.

4.5.1 Measurement units

Loss Function

For a classification model that provides raw scores for each class, Cross Entropy Loss is the appropriate loss function. Despite the fact that our samples will be appropriately balanced at each cycle, we want to punish melanoma loss by 50%.

Precision-Recall

Precision and recall are performance measurements used in pattern recognition, information retrieval, and classification (machine learning) that apply to data recovered from a collection, corpus, or sample space.

P=Div {TP, TP+FP}

R=Div {TP, TP+FN}

P-Precision

R-Recall

TP-True Positive

FP-False Positive

FN-False Negative

AUC ROC

ROC (Receiver Operating Characteristic) curve and AUC (Area Under the Curve) are commonly used evaluation metrics in machine learning, especially for binary classification problems.

ROC curve is a graphical representation of the performance of a binary classifier system as its discrimination threshold is varied. The curve is created by plotting the True Positive Rate (TPR) against the False Positive Rate (FPR) at various threshold settings. The TPR (also known as recall or sensitivity) is the proportion of actual positive cases that are correctly identified as positive by the classifier. The FPR is the proportion of negative cases that are incorrectly classified as positive.

AUC is a numerical value that represents the overall performance of a classifier system based on its ROC curve. The AUC value ranges from 0 to 1, where an AUC of 1 indicates a perfect classifier, while an AUC of 0.5 represents a random classifier that is no better than chance. The AUC value can be interpreted as the probability that a randomly chosen positive example is ranked higher by the classifier than a randomly chosen negative example.

4.6 Deployment

The final step is to deploy the ML model in a clinical setting to assist dermatologists in diagnosing skin lesions accurately.