Interpretable Skin Cancer Detection via Optimized CNN Models for Smart Healthcare Solutions

Abstract

Skin cancer is a common and potentially life-threatening condition, highlighting the importance of reliable and efficient diagnostic techniques. Recently, convolutional neural networks (CNNs) have demonstrated significant potential in automating the classification of skin cancer using thermoscopic images. Despite these advancements, the lack of interpretability in these models poses a barrier to their widespread use in clinical settings. In this study, we propose an interpretable CNN architecture optimized for skin cancer classification within a smart healthcare system. Our proposed model leverages a combination of traditional CNN layers with attention mechanisms and feature visualization techniques to enhance interpretability. Through extensive experimentation and hyperparameter tuning, we optimize the architecture to achieve both high accuracy and interpretability. We train and evaluate the model on a large dataset of thermoscopic images, ensuring its robustness across different skin cancer types and variations. Furthermore, we integrate the optimized CNN model into a smart healthcare system, allowing for seamless integration into clinical workflows. The system offers real-time feedback to dermatologists, helping them make better decisions and enhancing the accuracy of their diagnoses. Moreover, the interpretable nature of our model enables clinicians to understand the underlying features driving classification decisions, fostering trust and acceptance. Our experimental results demonstrate superior performance compared to state-of-the-art methods, achieving high accuracy in skin cancer classification while providing meaningful insights into the decision-making process. Overall, our proposed approach represents a significant advancement toward the development of interpretable and effective tools for skin cancer diagnosis in smart healthcare systems