State of the Art of Skin Anomaly Detection Using Machine Learning Techniques

Abstract

Skin cancer is one of the most common and deadly forms of cancer, for which early detection remains critical. In this paper, we provide a comprehensive review and technical comparison of artificial intelligence (AI) techniques for skin lesion analysis.We discuss a wide variety of conventional machine learning and latest deep learning techniques, with particular focus on pre-trained convolutional neural networks (CNNs) such as VGG, ResNet, EfficientNet, and MobileNet. An in-depth comparative analysis is provided in the form of classification accuracy, model complexity, training time, and inference speed. Preprocessing techniques—like hair removal, image normalization, artifact removal, and data augmentation techniques like geometric and color transformations—are examined for their effectiveness in enhancing robustness, particularly on real-world datasets like HAM10000, ISIC, and PH2. In addition, the paper addresses the feasibility of executing these models on embedded hardware platforms (Raspberry Pi, Jetson Nano, Coral Edge TPU) by analyzing their TensorFlow Lite and ONNX support and the impact of model compression techniques (quantization, pruning, distillation). The final cross-comparison identifies the optimal combination of dataset, model architecture, and hardware platform for creating an inexpensive, interpretable, and real-time skin lesion detection system. The paper concludes with the key takeaways for prototyping an embedded diagnostic system for low-resource environments.