Hybrid Skin Lesion Detection Integrating CNN and XGBoost for Accurate Diagnosis

Abstract

Skin cancer, particularly melanoma, remains one of the most challenging medical conditions due to its rapid progression and high mortality rate when not detected early. The growing prevalence of skin cancer highlights a significant problem in medical diagnostics: the need for automated, accurate, and efficient classification systems that can aid dermatologists in diagnosing various types of skin lesions. This issue is exacerbated by the imbalance in available datasets, underrepresentation of certain lesion classes, and a lack of generalizable diagnostic tools, ultimately impacting patient outcomes and healthcare efficiency.

This study aimed to develop and evaluate a hybrid model integrating Convolutional Neural Networks (CNNs) for feature extraction and XGBoost for classification to address the problem of skin lesion classification. This study's guiding conceptual framework was applying deep learning techniques combined with ensemble models to enhance classification accuracy and model interpretability.

The study utilized the HAM10000 dataset, comprising 10,015 dermatoscopic images across seven skin lesion classes. Dynamic resampling based on power analysis ensured class balance by selecting 158 samples per class. Image preprocessing techniques, such as resizing, hair removal, and Gaussian blurring, were applied to standardize the data. The CNN model extracted hierarchical features, while the XGBoost model performed classification on these features. The research methodology involved a quantitative approach using performance metrics such as accuracy, precision, recall, F1-score, and ROC-AUC to evaluate the model’s effectiveness.

The results demonstrated that the CNN-XGBoost hybrid model achieved superior classification performance with an accuracy of 86.46% on the test dataset, outperforming the standalone CNN model. The hybrid model effectively addressed class imbalance and exhibited high discriminatory power across all lesion classes, as confirmed by an average ROC-AUC score of 0.98.

The study concludes that the hybrid CNN-XGBoost model holds significant potential for assisting dermatologists in early skin lesion detection and improving diagnostic accuracy. Recommendations for future research include validation using diverse datasets, incorporating clinical metadata, and enhancing model interpretability for real-world deployment. These findings contribute to advancing AI-driven healthcare solutions, offering promising implications for dermatological diagnostics and patient care.