Multimodal Deep Learning for Lymph Node Metastasis in Thyroid Cancer

Area Under the Receiver Operating Characteristic Curve for a Multimodal Deep Learning Model Based on Cervical Ultrasound in Predicting Lymph Node Metastasis

The researcher will employ a multimodal deep learning model that integrates preoperative cervical color Doppler ultrasound images with corresponding structured text reports. The final output of the model is a predicted probability of lymph node metastasis for each patient (a continuous value between 0 and 1). This predicted probability will be compared with postoperative histopathological diagnosis results (the gold standard). A receiver operating characteristic curve will be plotted for the model, and its area under the curve will be calculated.This is the gold standard metric for evaluating the discriminative ability of a binary classification model (metastasis vs. non-metastasis). A higher AUC value indicates stronger overall discriminative power of the model.

Sensitivity of a Multimodal Deep Learning Model Based on Cervical Ultrasound for Predicting Lymph Node Metastasis

This metric aims to evaluate the ability of the constructed multimodal deep learning model to correctly identify patients with papillary thyroid carcinoma who truly have cervical lymph node metastasis, under the optimal diagnostic threshold. Researchers need to collect the number of patients diagnosed with lymph node metastasis through postoperative pathology, as well as the number of patients predicted as "positive" (i.e., predicted to have metastasis) by the model, in order to calculate the sensitivity of the cervical ultrasound-based multimodal deep learning model in predicting lymph node metastasis. Calculation formula: Sensitivity = Number of true positive patients / Total number of positive patients confirmed by postoperative pathology.

Specificity of a Multimodal Deep Learning Model Based on Cervical Ultrasound for Predicting Lymph Node Metastasis

This metric aims to evaluate the ability of the constructed multimodal deep learning model to correctly rule out patients with papillary thyroid carcinoma who have not developed cervical lymph node metastasis, under the optimal diagnostic threshold. Researchers need to collect the number of patients diagnosed without lymph node metastasis via postoperative pathology, as well as the number of patients predicted by the model as "negative" (i.e., predicted to have no metastasis), in order to calculate the specificity of the cervical ultrasound-based multimodal deep learning model in predicting lymph node metastasis. Calculation formula: Specificity = Number of true negative patients / Total number of negative patients confirmed by postoperative pathology.

The pathologically confirmed lymph node metastasis rate in the study cohort

It refers to the percentage of patients with at least one metastatic lymph node confirmed by postoperative pathological examination, relative to the total number of individuals in the corresponding study population. Researchers need to collect the number of patients diagnosed with lymph node metastasis through postoperative pathological examination.

Adjusted Odds Ratios for Clinical Factors Associated with Pathologically Confirmed Lymph Node Metastasis

Researchers need to collect the outcome variable (i.e., postoperatively pathologically confirmed lymph node metastasis status) and its exposure variables (such as the specific location of the primary tumor within the thyroid gland, maximum tumor diameter, patient age, etc.). Using these variables, the adjusted odds ratios are calculated to reflect, after adjusting for other confounding factors, how many times more likely patients with a specific exposure characteristic (e.g., tumor located in the upper pole) are to have lymph node metastasis compared to patients in the reference group (e.g., tumor located in the lower pole).

The weighted Kappa coefficient for the consistency between model-predicted pTNM stage and pathological stage

Researchers need to collect and record the model-predicted pTNM stage and the patient's true pTNM stage to evaluate the consistency between the model-predicted complete pTNM stage and the pathological stage.