Detection of Lymph Node Metastasis in Thyroid Cancer Using Deep Learning and Second Harmonic Generation Imaging.

Abstract

Papillary thyroid carcinoma (PTC) is the most prevalent type of thyroid cancer, with a significant proportion of patients being susceptible to lymph node metastasis (LNM). The presence of LNM has been shown to accelerate tumor progression. Existing diagnostic approaches, such as ultrasonography and postoperative pathological analysis, exhibit limited sensitivity in detecting non-metastatic cases, thus undermining subsequent treatment planning. In this investigation, an innovative automated quantitative histological classification framework called the Automatic Thyroid Cancer Lymph Node Metastasis Classification Network (AutoThyroLNMNet) is introduced, which integrates Second-harmonic generation (SHG) imaging technology with deep learning to detect LNM in thyroid cancer. A combined model was constructed utilizing the Pyramid Vision Transformer v2 (PVTv2) as the backbone of the deep learning architecture and incorporating a multi-layer perceptron to fuse deep learning outputs, pathological information, and the two categories of collagen features. The models demonstrated a strong performance on training sets, with the highest efficacy achieved for the model that incorporated 3D texture features, achieving an area under the receiver operating characteristic (ROC) curve of 0.99. These results suggest that AutoThyroLNMNet is capable of automatically and quantitatively classifying lymph node metastasis in thyroid cancer, offering a novel and effective tool for the precise detection of LNM.