Automatic diagnosis for thyroid nodules in ultrasound images by deep neural networks

Thyroid cancer is a disease in which the first symptom is a nodule in the thyroid region of the neck. It is one of the cancers with the highest incidences, and has the highest increase rate in the last thirty years. Ultrasonography is one of the most sensitive and widely used methods for detecting thyroid nodules. To assist in the analysis of thyroid ultrasound images, many computer-aided diagnosis methods have been proposed. Most of these methods perform diagnosis using only a single ultrasound image instead of using all images from an examination, which loses the overall information related to the thyroid nodules. However, in an ultrasound examination, the sonographer analyzes the thyroid nodule based on multiple images from different views. In the current study, a deep learning method is proposed to diagnose thyroid nodules using multiple ultrasound images in an examination as input. An attention-based feature aggregation network is proposed to automatically integrate the features extracted from multiple images in one examination, utilizing different views of the nodules to improve the performance of recognizing malignant nodules in the ultrasound images. To train and evaluate the proposed method, a large dataset is constructed. The experimental results demonstrate that our method achieves comparable performance with state-of-the-art methods for the diagnosis of thyroid ultrasound images.