PTC-CapsNet: capsule network for papillary thyroid carcinoma pathological images classification

Automatic classification of pathological images is an important task in the thyroid carcinoma histopathological analysis. Currently, histological diagnosis has become a leading field in medical imaging computing, which can reduce the burden of pathologists and the misdiagnosis rate. At present, works on designing convocational neural networks(CNNs) for automatically classifying pathological images are increasing rapidly. However, classification of thyroid carcinoma pathological images remains challenging due to the small differences between benign and malignant observed from pathological images. The existing CNN methods cannot classify the pathological images well. Inspired by the diagnosis process observed by the pathologists and the reasoning ability of CapsNet, we propose a novel capsule network named PTC-CapsNet based on semantic features derived from multi-magnification pathological images for papillary thyroid carcinoma (PTC) pathological images classification. The proposed method is more in line with the diagnostic decision and visual cognitive of pathologist. It can deal with spatial hierarchies between complex pathological features excellently. A new loss function named penalty binary cross-entropy is designed to reduce the false negative rate. Furthermore, an eye movement dataset is established by pathologists to evaluate the performance of models subjectively, while a new metric also is proposed to verify the performance of these methods objectively. The experimental results illustrate the superiority of our method on our PTC pathological images dataset and the BreakHis dataset.