An automatically thyroid nodules feature extraction and description network for ultrasound images

Thyroid nodules are mainly detected and diagnosed using ultrasound (US) images. However, interpreting and summarizing the content of thyroid US images is a time-consuming task event for experienced radiologist. Thus, an intelligent method that can complete the mapping from visual information to concise description is very necessary. A deep learning network structure for thyroid nodules features description in US images is proposed. This network consists of 2 parts: 1. the visual features extraction part; 2. the language description generation part. The proposed framework can automatically extract the features of the thyroid nodules and generate feature description reports based on the input ultrasound image and its segmentation results. A description accuracy of 95% is achieved using clinical US images.