Comparative Analysis of Diverse Activation Functions and Kernel Methods for Thyroid Cancer Classification

Thyroid carcinoma ranks among the most common malignancies affecting the endocrine system, and early diagnosis through histopathological image analysis is crucial for improving patient outcomes. Despite advances in computeraided diagnostics, challenges persist in optimizing model architectures and configurations for complex histopathological patterns. Addressing this gap, the present research proposes a comprehensive deep learning framework for thyroid cancer classification, leveraging a novel, expert-annotated histopathological dataset curated at Sultan Qaboos University Hospital (SQUH). We systematically compare the performance of Convolutional Neural Networks (CNNs) and Support Vector Machines (SVMs), evaluating multiple activation functions (ReLU, sigmoid, tanh, identity) and kernel methods (linear, polynomial, RBF, sigmoid). The framework emphasizes how activation and kernel selection critically influence feature representation, learning stability, and generalization ability in histopathological analysis. Experimental results demonstrate that CNNs, particularly ResNet50 with ReLU activation, achieve superior diagnostic accuracy, sensitivity, and F1-scores compared to SVMs, with RBF kernel SVMs showing competitive but lower performance. By integrating advanced network designs with rigorous kernel optimization, this study contributes significantly to the development of robust, scalable AI-based diagnostic systems for thyroid cancer. Furthermore, the release of a high-resolution, pathologist-verified dataset offers a vital resource for future research in medical image classification and AI-assisted pathology. This research provides perceptions into the optimal machine learning configurations for thyroid cancer classification and contributes to the development of precise and scalable diagnostic tools for histopathological image analysis.