Artificial Intelligence-Based Skin Lesion Analysis and Skin Cancer Detection

Abstract

Advanced diagnostic methods are necessary for the early and precise diagnosis of skin cancer, a deadly disease that poses a danger. The accuracy of manual skin lesion assessment and visual inspection is limited, which is why sophisticated diagnostic tools are required. In response, this study presents a groundbreaking approach that makes use of an ensemble of twelve pre-trained deep learning models, including InceptionV3, VGG16, VGG19, Xception, DensNet121, DensNet201, ResNet152V2, MobileNet, MobileNetV2, ConvNeXtLarge, NASNetMobile, and InceptionResNetV2. This study demonstrates a distinct training strategy by employing a two-phase approach: first, training only the newly added dense layers while maintaining the layers of the base model frozen, and then, fine-tuning the entire model. This sophisticated process improves CNN convolutions' stability during feature extraction, which in turn improves the model's overall performance in terms of prediction accuracy. The HAM10000 dataset was used as the main basis for training, evaluating, and comparing all of the models used in this comprehensive research, assuring a consistent and exacting method to progress the field of skin cancer classification. The model with the highest classification accuracy, ResNet152V2, with an F1 score of 98%, wins. By recognizing the intricacy of skin lesions, the study makes the significance of its findings clear and provides hope for the development of more advanced diagnostic instruments. This article not only offers a critical assessment of current methods but also tackles problems and indicates future directions for future research in the field of medical image categorization. This research has implications that extend beyond skin cancer diagnosis; it impacts several therapeutic applications and provides a solid foundation for further advancements in the field.