Multimodal Nested Attention Network for Lymph Node Metastasis Prediction of Thyroid Carcinoma

Abstract

Accurate preoperative prediction of cervical Lymph Node Metastasis (LNM) is critical for surgical decision-making in thyroid cancer patients, and the difficulty in it often leads to over-treatment. UltraSound (US) and Computed Tomography (CT) are two primary non-invasive examinations, but neither method alone provides satisfactory diagnostic accuracy. To address this problem, we propose a Multimodal Nested Attention Network (MNANet) to integrate US and CT images. The network is designed to extract specific complementary information from US and CT images, and comprehensively fuse multimodal features at multiple granularities. In our internal cohort, MNANet achieves Areas Under the Curves (AUCs) of 0.88 and 0.86 for central and lateral cervical sites, respectively, representing a significant improvement of 0.06 to 0.10 compared to unimodal models and outperforming state-of-the-art medical multimodal methods across all metrics. The model demonstrates robust cross-institutional generalization and maintains superior performance across other imaging modalities(e.g., Magnetic Resonance Imaging (MRI)). Additionally, our model exhibits a more precise focus on the thyroid nodule, indicating enhanced learning ability. Moreover, we systematically evaluate the applicability across various clinical characteristics, identifying individuals who can benefit most from the multimodal approach.

Ähnliche Arbeiten