An Explainable Digital Twin Framework for Skin Cancer Analysis Using Early Activation Meta-Learner

Pradeepa Sampath¹S Gopika¹S Vimal²Joonje Kang³Sanghyun Seo^4*

• ¹Department of Information Technology, School of Computing, SASTRA Deemed University, Thanjavur, India
• ²Department of Artificial Intelligence and Data Science, Sri Eshwar College of Engineering, Coimbatore, India
• ³Department of Applied Art and Technology, Chung-Ang University - Da Vinci Campus, Anseong-si, Republic of Korea
• ⁴School of Art and Technology, Chung-Ang University - Da Vinci Campus, Anseong-si, Republic of Korea

Skin cancer is among the most common cancers globally, which calls for timely and precise diagnosis for successful therapy. Conventional methods of diagnosis, including dermoscopy and histopathology, are significantly dependent on expert judgment and therefore are time-consuming and susceptible to inconsistencies. Deep learning algorithms have shown potential in Skin Cancer classification but tend to consume huge amounts of computational resources and large training sets. To overcome these issues, we introduce a new hybrid computer-aided diagnosis (CAD) system that integrates Stem Block for feature extraction and machine learning for classification. The ISIC skin cancer dermoscopic images are collected from the Kaggle and the essential features are collected from the Stem Block in deep learning algorithm. The selected features, which are standardized using StandardScaler to transform the extracted features to have zero mean and unit variance, are then classified using a Meta-Learning Classifier to enhance precision and efficiency. Additionally, a Digital Twin framework is introduced to simulate and analyze the diagnostic process virtually, enabling real-time feedback and performance monitoring. This virtual replication aids in continuous improvement and supports the deployment of the CAD system in clinical environments.To improve transparency and clinical reliability, Explainable AI (XAI) methods are incorporated to visualize and interpret model predictions. Compared to state-of-the-art approaches, our system reduces training complexities without compromising the high classification precision. Our proposed model attains an accuracy level of 96.25%, which demonstrates its consistency and computationally efficient status as a screening tool for detecting skin cancer.