DEEP LEARNING APPROACHES FOR SKIN CANCER DIAGNOSIS FROM LESION IMAGING

Abstract

The current paper proposes the solution to the detection of early skin cancer using deep learning and using the ISIC 2018 dataset with specific emphasis on image processing and advanced model’s design. One of the health problems that are of great concern is skin cancer and especially melanoma, early identification of the disease is highly relevant when it comes to the survival of individuals. Although the traditional diagnostic methods are effective, these are, in most cases, limited by the skills and experience of the dermatologists and hence, the results do not always coincide. To this end, it is proposed in the paper to take advantage of the Convolutional Neural Networks (CNNs) or in this instance ResNet50, InceptionV3 and Inception ResNet, and image preprocessing techniques where normalization, augmentation and Super-Resolution using ESRGAN can be used to enhance the quality of the lesion images. The pre-trained models are fine-tuned on the task of skin cancer classification using transfer learning in the study. The models have been evaluated using such metrics as accuracy, precision, recall, F1-score, and Area Under Curve (AUC) of the Receiver Operating Characteristic (ROC) curve. The discoveries are that, Inception ResNet has the highest accuracy and the value is 87.6, 86.5, 88.1, 87.4 and the area under the curve is 0.91. InceptionV3 is performing quite well as well with accuracy of 86.2 and AUC of 0.89. ResNet50 achieves an accuracy of 85.3 percent and AUC of 0.88. According to CNN baseline model, its accuracy is reasonable at 84.5 per cent but it is also less reliable than the more complicated models. These findings indicate that deep learning can be highly effective in the process of identifying skin cancer and the best and most consistent model is Inception ResNet. The article is practical as it demonstrates how the already existing deep learning approaches can be as effective and even more effective in skin cancer diagnostics than a trained dermatologist and serve as a means of early and reliable skin cancer detectors.