ThyHisTer: A new thyroid histopathology image dataset for ternary classification of thyroid cancer

Thyroid cancer is a common type of endocrine cancer, and its incidence rate has been increasing year by year. Due to the scarcity of publicly accessible histopathology image datasets for thyroid cancer diagnosis, it is difficult to develop automatic Computer-aided Diagnostic (CAD) systems for enhancing the accuracy of thyroid cancer diagnosis. To address this issue, this work aims to construct a novel publicly accessible thyroid histopathology image dataset for ternary classification of thyroid cancer, namely ThyHisTer. Furthermore, to present a benchmarking performance evaluation on the ThyHisTer dataset, this work explores the performance of various deep learning methods on thyroid cancer classification tasks. Additionally, this work proposes a new lightweight deep learning model called SeSepViT for thyroid cancer classification, which integrates the advantages of Squeeze and Excitation (SE) networks and Separable Vision Transformer (SepViT). This work conducts extensive experiments on the collected ThyHisTer dataset, and utilize various deep learning methods to validate the performance of thyroid cancer classification. Experimental results show that the proposed SeSepViT achieves highly comparable performance to other used deep learning methods on thyroid cancer classification tasks, and simultaneously exhibits relatively lower computational cost. The release of ThyHisTer is expected to facilitate the application of advanced deep learning methods for automatic thyroid cancer diagnosis, thereby assisting doctors in early detecting thyroid cancer in clinical practice. The code is available on https://github.com/beatttt/ThyHisTer.