The challenge in the predictions of skin lesions is due to the noise and contrast. The manual dermoscopy imaging procedure results in the wrong prediction. A deep learning model assists in detection and classification. The structure in the proposed handles CNN architecture with the stack of separate layers that use a differential function to transform an input volume into an output volume. For image recognition and classification, CNN is specifically powerful. The model was trained using labeled data with the appropriate class. CNN studies the relationship between input features and class labels. For model building, use Keras for front-end development and Tensor Flow for back-end development. The first step is to pre-process the ISIC2019 dataset, splitting it into 80% training data and 20% test data. After the training and test splits are complete, the dataset has been given to the CNN model for evaluation, and the accuracy on each lesion class was calculated using performance metrics. The comparative analysis has been done on pretrained models like VGG19, VGG16, and MobileNet.
TopIntroduction
Melanoma is the type of skin cancer that begins to grow out of control of the development of melanocytes (PS Staff, 2016). In this regard, the main factors for detecting skin cancer and distinguishing between benign and melanoma, such as symmetry, colour, size and shape (PS Staff, 2016). Many countries worldwide, especially the United States, report growing death rates from skin cancer (Marks, 1995). Recent cancer statistics and figures show that the calculable range of recent cancer cases of this kind is around 1.9 million and that in the United States the death rate will be around 608,570 (Siegel et al., 2023). Earlier diagnosis is likely to reduce the death rate. Daylight exposure is associated with the greatest risk of carcinoma development with every malignant melanoma and non-melanoma cancer in the skin. Current carcinoma encompasses malignant melanoma and NMSC malignancies, made up of basal (BCC) and squamous cell carcinoma (SCC), as indicated by Figure1 (Gordon, 2013). Melanoma is the deadliest type of cancer occurring in human beings that leads to coloured markings or skin moles. Clinical testing, dermoscopic image analysis, histological investigation, and ultimate biopsy are the initial diagnosis of carcinoma (Mane & Shinde, 2018).
Skin lesion classes:
Lesions images of skin are classified into seven classes.
Classification of Lesion images of Skin:
TopDeep learning architectures are now used in medical image analysis and new frameworks are being developed to predict, diagnose and detect. Current neural networks that are good for image identification and outperform the categorization of skin cancer (Adegun & Viriri, 2020). The first job of classifying the image is to accept the image and how it is characterised by its class. The ability to recognise human images is quite different from the machine. The machine detects the images as pixels. CNN’s are neural networks specially built for the detection of images and computer vision challenges. Unlike Machine Learning techniques, CNN uses a sequence of convolutionary, pooling and non-linear layer to process and pre-process the picture as a 2D vector and ultimately FCN (fully connected layer) to produce the output (Yamashita et al., 2018; Aghdam et al., 2017).
CNN-based profound learning algorithms have demonstrated amazing detection, classification and segmentation performance in medical imaging applications (Harley, 2015).
Zhang (2021) suggested a strategy for retrieving deep skin injury characteristics using deep CNNs. Models such as AlexNet, ResNet-18 and VGG 16 have been pre-trained. In the last stage, the features generated are provided to the SVM classification for training. Classifier used for classification purposes. The model was assessed using the 2017 ISIC dataset and the accuracy achieved was 97.55%.
Dorj et al. (2018) introduced the method of categorization by pretrained deep CNN of various types of skin lesion images (AlexNet used for feature extraction). The system provided produced the highest mean sensitivity, specificities, SCC precision, actinic keratosis (AK), and BCC values: 95.1% (98.9%), and 94.17% respectively.