A robust CNN classification of whole slide thyroid carcinoma images

Abstract

The objective of this paper is to build a classification system for "Whole Slide Images" (WSIs) based on a Convolutional Neural Network (CNN). Six types of thyroid tumors can be classified by the system: "follicular adenoma" (FA), "papillary carcinoma" (PC), "follicular carcinoma" (FC), "papillary follicular variant" (PFV), "poorly-differentiated follicular carcinoma" (PDFC), and "well-differentiated follicular carcinoma" (WDFC). The proposed custom CNN is compared with the well-known pre-trained Alexnet CNN. The results show the robustness of the proposed CNN, achieving an overall accuracy of 97.07% compared to only 93.81% for the Alexnet.