Abstract: PURPOSE: Recently, deep learning has been increasingly applied in medical image analysis. Significant results have been achieved in tasks such as classification and segmentation of tumors. Unfortunately, there are two main difficulties in this process. One is that it is time-consuming and subjective to manually delineate the tumor edge. The second is that the unexplained and non-robust nature of deep learning makes doctors feel uncomfortable when using them clinically.
METHODS: To solve the above problems, we propose an attention-based large receptive field network (ALRN) and apply it to the benign and malignant classification tasks of thyroid nodules of the ultrasound image. The network is divided into two parts: the trunk channel and the attention channel. The input of the trunk channel is an ultrasound image of the thyroid gland, which is mainly used to adaptively learn the texture, shape and other features of the nodule. The input of attention channel is a nodule position template for the doctor to rough sketch, which is used as a priori knowledge of attention. It adaptively learns the area that the network needs to be concerned with and constrains the trunk channel.
RESULTS: The ALRN has been tested to have better accuracy and robustness than normal convolutional neural networks (CNNs). Furthermore, the area of attention is better matched to the clinician's experience. It is worth mentioning that because the network has a large receptive field, the doctor does not need to carefully depict the edge of the tumor, and only needs to roughly sketch the position of the nodule.
CONCLUSIONS: We propose a large receptive convolutional neural network based on the attention mechanism and achieve good results in the task of benign and malignant classification of thyroid nodules of the ultrasound image.

Key words: Deep learning, attention mechanism, thyroid nodule, ultrasound image