A LightGBM-Based Diagnostic Model for Skin Cancer Classification Using Dermoscopic Images

Skin cancer is one of the most common types of cancer across the globe, and detecting it early with high accuracy is vital for better patient outcomes. This paper introduces a machine learning approach designed to classify various types of skin lesions by analyzing dermoscopic images along with relevant clinical information. The proposed method employs Light Gradient Boosting Machine (LightGBM), a robust and efficient ensemble method well-suited for handling class imbalance and high-dimensional feature spaces. In the development of automated skin lesion classifiers, the HAM10000 collection—consisting of more than then thousand dermoscopy images and associated metadata including patient age, sex, and lesion location—is leveraged by extracting high-dimensional features from each image using a pretrained ResNet18 convolutional neural network. These image features are then fused together with the patient metadata, yielding a holistic feature representation that captures both phenotypic appearance and relevant clinical factors for subsequent model trainingTo enhance generalization, we incorporate classweighted training, dimensionality reduction using Principal Component Analysis (PCA), and image augmentation techniques. Based on experimental findings, the model reaches an accuracy of 80%. These outcomes highlight the potential of integrating deep learning-based feature extraction with gradient boosting classifiers for automated skin cancer diagnosis. The proposed framework is interpretable, scalable, and accurate, making it well-suited for integration into clinical decision support systems.