Hyper-net: A Dual-Path Convolutional Network for Thyroid Fine Needle Aspiration Cell Segmentation

In recent years, hundreds of thousands of people around the world have been diagnosed with thyroid cancer, and this number is increasing year by year. Thyroid fine needle aspiration is the most sensitive and specific method to determine the nature of nodules before operation. The diagnosis of thyroid FNA needs to judge the benign and malignant cells on the smear, which largely depends on the doctor's experience and is quite challenging and time-consuming. Although some methods have good performance at present, it is not reliable to analyze FNA only by morphological classification. This paper collected the first hyperspectral pathological image data set of thyroid papillary carcinoma, which contains 117 hyperspectral images with pixel level labels of benign and malignant cells. It can be used as a benchmark for hyperspectral image analysis of thyroid papillary carcinoma. Irregularity and overlap are the prominent features of pathological images of papillary thyroid carcinoma, which brings challenges to accurate segmentation. Based on the collected data set, we designed a dual path network structure (Hyper-net) with attention mechanism for cell segmentation. The segmentation network combines the attention mechanism, spatial and spectral information to effectively complete the segmentation task in thyroid FNA smears. The global class attention module can extract the key features of each class in the whole data set and enhance the global perception of the network. Compared with the benchmark network and other dual path networks, our hyper-net with attention mechanism achieves much better performance for thyroid fine needle aspiration cell segmentation.