Diagnosis of cervical lymph node metastasis with thyroid carcinoma by deep learning application to CT images.

Introduction: The incidence of thyroid diseases has increased in recent years, and cervical lymph node metastasis (LNM) is considered an important risk factor for locoregional recurrence. This study aims to develop a deep learning-based computer-aided diagnosis (CAD) method to diagnose cervical LNM with thyroid carcinoma on computed tomography (CT) images.

Methods: A new deep learning framework guided by the analysis of CT data for automated detection and classification of LNs on CT images is proposed. The presented CAD system consists of two stages. First, an improved region-based detection network is designed to learn pyramidal features for detecting small nodes at different feature scales. The region proposals are constrained by the prior knowledge of the size and shape distributions of real nodes. Then, a residual network with an attention module is proposed to perform the classification of LNs. The attention module helps to classify LNs in the fine-grained domain, improving the whole classification network performance.

Results: A total of 574 axial CT images (including 676 lymph nodes: 103 benign and 573 malignant lymph nodes) were retrieved from 196 patients who underwent CT for surgical planning. For detection, the data set was randomly subdivided into a training set (70%) and a testing set (30%), where each CT image was expanded to 20 images by rotation, mirror image, changing brightness, and Gaussian noise. The extended data set included 11,480 CT images. The proposed detection method outperformed three other detection architectures (average precision of 80.3%). For classification, ROI of lymph node metastasis labeled by radiologists were used to train the classification network. The 676 lymph nodes were randomly divided into 70% of the training set (73 benign and 401 malignant lymph nodes) and 30% of the test set (30 benign and 172 malignant lymph nodes). The classification method showed superior performance over other state-of-the-art methods with an accuracy of 96%, true positive and negative rates of 98.8 and 80%, respectively. It outperformed radiologists with an area under the curve of 0.894.

Discussion: The extensive experiments verify the high efficiency of the proposed method. It is considered instrumental in a clinical setting to diagnose cervical LNM with thyroid carcinoma using preoperative CT images. The future research can consider adding radiologists' experience and domain knowledge into the deep-learning based CAD method to make it more clinically significant.

Conclusion: The extensive experiments verify the high efficiency of the proposed method. It is considered instrumental in a clinical setting to diagnose cervical LNM with thyroid carcinoma using preoperative CT images.