Radiomics and deep learning for large volume lymph node metastasis in papillary thyroid carcinoma

Area of Science:

• Oncology
• Radiology
• Artificial Intelligence

Background:

• Thyroid cancer frequently involves early lymph node metastasis (LNM).
• Large volume LNM (LVLNM) is associated with poorer patient prognosis.
• Accurate pre-operative prediction of LVLNM is crucial for treatment planning.

Purpose of the Study:

• To develop and validate a predictive model for LVLNM in papillary thyroid carcinoma (PTC) patients.
• To integrate radiomics and deep learning (DL) features for enhanced prediction accuracy.
• To assess the model's performance in both internal and external validation datasets.

Main Methods:

• A multicenter retrospective study of 854 PTC patients.
• Extraction of 1,357 radiomics features and application of various machine learning algorithms (LR, SVM, KNN, MLP, RF, ExtraTrees, XGBoost, LightGBM).
• Development of DL models using AlexNet, DenseNet121, inception_v3, ResNet50, and transformer architectures.
• Creation of a combined radiomics-DL model (Thy-DL-Radiomics) and evaluation using AUC, ACC, SEN, SPE, PPV, NPV, and F1-score.

Main Results:

• The ExtraTrees radiomics model achieved an AUC of 0.787; the DenseNet121 DL model achieved an AUC of 0.766.
• The combined Thy-DL-Radiomics model demonstrated superior performance with an AUC of 0.839 (internal validation) and 0.789 (external validation).
• LASSO identified key radiomics and metabolite features for model optimization.

Conclusions:

• An integrated radiomics and deep learning model effectively predicts LVLNM in PTC patients.
• This predictive capability can inform personalized treatment strategies for thyroid cancer.
• The Thy-DL-Radiomics model offers a promising tool for pre-operative risk stratification.