# Multimodal ultrasound radiomics containing microflow images for the prediction of central lymph node metastasis in papillary thyroid carcinoma

## TL;DR

This study uses ultrasound radiomics and machine learning to predict lymph node metastasis in thyroid cancer patients, improving preoperative decision-making.

## Contribution

The novel use of microflow images in a multimodal radiomics model for predicting central lymph node metastasis in papillary thyroid carcinoma.

## Key findings

- A fusion model using clinical data and multimodal ultrasound radiomics achieved high diagnostic performance (AUC = 0.886 in training, 0.873 in testing).
- Microflow images contributed significantly to the model, confirming their diagnostic value for central lymph node metastasis.
- The model provides a net clinical benefit for preoperative evaluation of lymph node status in papillary thyroid cancer patients.

## Abstract

This study aimed to construct a model by applying radiomics and machine learning (ML) to multimodal ultrasound images (including grayscale, elastography and microflow images) along with clinical data to predict central lymph node metastasis (CLNM) in patients with papillary thyroid cancer (PTC).

A cohort of 213 patients who underwent thyroidectomy accompanied by lymph node dissection (LND) and were pathologically diagnosed with PTC postoperatively was enrolled and randomized to the training cohort (n = 170) or testing cohort (n = 43). Radiomics features were extracted from multimodal images and subsequently screened via the least absolute shrinkage and selection operator (LASSO). The same methods were applied to screen clinical features. Nine ML algorithms were used to construct clinical models, radiomics models and fusion models. Model performance was assessed via receiver operating characteristic curves (ROC), decision curve analysis (DCA), and Delong test. Finally, the optimal model was interpreted and visualized via Shapley additive explanation (SHAP).

In each modality, 1561 features were extracted from the ultrasound images. Sixteen features were ultimately retained, including 6 grayscale features, 6 elastography features, and 4 microflow features. From the clinical features, including gender, age, traditional ultrasound signs and serological indicators, 2 relevant features were selected. Among the prediction models, the fusion model constructed by Multilayer Perceptron (MLP) algorithm showed the best diagnostic performance, outperforming the other models in both the training cohort (AUC = 0.886) and the testing cohort (AUC = 0.873).

The fusion model based on clinical data and multimodal ultrasound radiomics has better predictive ability and net clinical benefit for CLNM in patients with PTC, confirms the diagnostic value of microflow images for CLNM, and can help to evaluate patients’ preoperative lymph node status and make the correct decision on the surgical procedure.