Classification of Thyroid Carcinoma in Whole Slide Images Using Cascaded CNN
Ahmed S. El-Hossiny, Walid Al‐Atabany, Osama Hassan, Ahmed M. Soliman, Sherif A. Sami
Abstract
The objective of this research is to build a “Whole Slide Images” classification system using Convolutional Neural Network (CNN). This system is capable of classifying Thyroid tumors into three types: Follicular adenoma, follicular carcinoma, and papillary carcinoma. Furthermore, the cascaded CNN technique is additionally employed to classify the classified follicular carcinoma into four subclasses: follicular carcinoma, papillary follicular variant, well-differentiated follicular carcinoma, and Poorly-differentiated follicular carcinoma. Results of the proposed CNN architecture showed effective classification of Thyroid carcinoma in the whole slide images with an overall accuracy of 94.69%. In the first classification stage, the images are classified into either one of three main types with an overall accuracy of 98.74%, while in the second classification stage, using the cascaded CNN, accuracy was 95.90% for further sub-classification into four sub-classes. Our cascaded CNN outperformed the accuracy of other studies due to splitting classification process of the thyroid into two stages which reduces the number of classes in each stage.