Histopathological Analysis of Fine-Needle Aspiration Biopsies of Thyroid Nodules Using Explainable Convolutional Neural Networks

Abstract

Thyroid Cancer is a disease in which abnormal cells grow uncontrollably in the gland with the potential to invade other organs. Every year, almost 44,000 new cases are diagnosed worldwide. Histopathological diagnosis of fine-needle aspiration biopsies of thyroid nodules is the most precise exam to confirm the diagnosis and estimate the stages of the disease. The diagnostic process in such an exam involves detecting atypical signs, such as the presence of cell proliferation with irregular shape and texture. This task could be even harder once you consider that most thyroid biopsies might present multiple pathological states, such as inflammatory diseases and hyperplasia. Therefore, this paper addresses the development of a Computer Vision method to assist the histopathological diagnosis of normal, thyroid papillary carcinoma and goiter. The proposed method model and implement a Convolutional Neural Network to detect visual patterns to differentiate the three pathological states. Experiments following the Holdout Cross-Validation protocol reached an accuracy of 88.73% for the multiclass approach and 95.74% accuracy for the binary assessment. The results confirm the potential of the proposed method to assist pathologists in prescribing a more precise diagnosis.