Deep Learning-Assisted SERS of Fine-Needle Aspiration Washout Fluid for Preoperative Papillary Thyroid Carcinoma Diagnosis and Lymph Node Metastasis Prediction
- 1Department of Thyroid Surgery, China-Japan Union Hospital of Jilin University, Changchun, Jilin, People’s Republic of China
- 2School of Information Engineering, Jiangxi Provincial Key Laboratory of Advanced Signal Processing and Intelligent Communications, Nanchang University, Nanchang, Jiangxi, People’s Republic of China
- 3State Key Laboratory of Advanced Manufacturing for Optical Systems, Changchun Institute of Optics, Fine Mechanics and Physics, Chinese Academy of Sciences, Changchun, Jilin, People’s Republic of China
Abstract
Background: Accurate preoperative assessment of papillary thyroid carcinoma (PTC) and cervical lymph node metastasis (LNM) is important. This study aimed to analyze molecular fingerprint in fine-needle aspiration (FNA) washout fluid using surface-enhanced Raman spectroscopy (SERS) and develop diagnostic models for nodule malignancy differentiation and LNM prediction integrating machine learning (ML) algorithms.
Methods: A retrospective analysis was conducted on 444 FNA washout fluid samples. AgNPs served as the SERS substrate. Following spectral preprocessing, prominent differential peaks were statistically screened and subjected to biomolecular assignments. The dataset was partitioned into training, validation, and test sets (6:2:2) to construct five predictive models: SVM, RF, LightGBM, XGBoost, and CNN. Diagnostic performance was evaluated based on the area under the receiver operating characteristic curve (AUC), accuracy, precision, NPV, recall and F1-score.
Results: Multiple significant differential peaks were identified, primarily reflecting alterations in nucleic acids, proteins, lipids, and the tryptophan-kynurenine metabolic pathway. For PTC diagnosis, PTCSpecNet model yielded a test set AUC of 0.991, cross-validation accuracy of 0.917, and recall of 0.979, while SVM achieved the highest test AUC (0.995) and LightGBM achieved the highest accuracy (0.978). For LNM prediction, LightGBM yielded the highest test AUC (0.916) and recall (0.909), whereas LymSpecNet model demonstrated competitive discriminative capacity (test set AUC = 0.913, cross-validation accuracy = 0.895) with an optimal clinical balance between precision (0.857) and recall (0.818). The fitting times for PTCSpecNet and LymSpecNet were 5.97 s and 8.52 s, respectively, while the estimated realistic clinical turnaround time is approximately 20–25 min, supporting rapid diagnostic feasibility.
Conclusion: SERS analysis of FNA washout fluid combined with deep learning effectively captures the molecular fingerprint of the in situ thyroid nodule microenvironment, enabling high-sensitivity diagnosis of PTC and reliable prediction of LNM. Future prospective multi-center studies with external validation cohorts are warranted to further establish its broad clinical generalizability.
Related
- Machine learning-based prediction of lymph node metastasis in papillary thyroid carcinoma: an interpretable multidimensional model
- High-volume cervical lymph node metastasis in papillary thyroid carcinoma: development and external validation of an interpretable prediction model.
- Development and Patient-Level Internal Validation of an MRI-Based Radiomics Machine Learning Model for Preoperative Prediction of Cervical Lymph Node Metastasis in Oral Squamous Cell Carcinoma: A Pilot Study.
- Treatment strategies for unilateral papillary thyroid carcinoma with ipsilateral cervical lymph node metastasis: a narrative review
- The clinical value of ultrasound radiomics combined with clinical data nomogram in predicting cervical lymph node metastasis of papillary thyroid carcinoma.
BibTeX
@article{he2026deep,
title = {{Deep Learning-Assisted SERS of Fine-Needle Aspiration Washout Fluid for Preoperative Papillary Thyroid Carcinoma Diagnosis and Lymph Node Metastasis Prediction}},
author = {Rundong He and Ping Sun and Le Zhou and Danheng Gao and Rui Du and Meng Luo and Chong Chen and Yiming Guo and Qihan Zhang and Hui Sun and Haoran Meng},
journal = {International Journal of Nanomedicine},
year = {2026},
month = oct,
doi = {10.2147/ijn.s632462},
url = {https://doi.org/10.2147/ijn.s632462},
note = {CC BY-NC 4.0},
}
RIS
TY - JOUR
TI - Deep Learning-Assisted SERS of Fine-Needle Aspiration Washout Fluid for Preoperative Papillary Thyroid Carcinoma Diagnosis and Lymph Node Metastasis Prediction
AU - Rundong He
AU - Ping Sun
AU - Le Zhou
AU - Danheng Gao
AU - Rui Du
AU - Meng Luo
AU - Chong Chen
AU - Yiming Guo
AU - Qihan Zhang
AU - Hui Sun
AU - Haoran Meng
PY - 2026
DA - 2026/10/30
JO - International Journal of Nanomedicine
DO - 10.2147/ijn.s632462
UR - https://doi.org/10.2147/ijn.s632462
AB - Background: Accurate preoperative assessment of papillary thyroid carcinoma (PTC) and cervical lymph node metastasis (LNM) is important. This study aimed to analyze molecular fingerprint in fine-needle aspiration (FNA) washout fluid using surface-enhanced Raman spectroscopy (SERS) and develop diagnostic models for nodule malignancy differentiation and LNM prediction integrating machine learning (ML) algorithms. Methods: A retrospective analysis was conducted on 444 FNA washout fluid samples. AgNPs served as the SERS substrate. Following spectral preprocessing, prominent differential peaks were statistically screened and subjected to biomolecular assignments. The dataset was partitioned into training, validation, and test sets (6:2:2) to construct five predictive models: SVM, RF, LightGBM, XGBoost, and CNN. Diagnostic performance was evaluated based on the area under the receiver operating characteristic curve (AUC), accuracy, precision, NPV, recall and F1-score. Results: Multiple significant differential peaks were identified, primarily reflecting alterations in nucleic acids, proteins, lipids, and the tryptophan-kynurenine metabolic pathway. For PTC diagnosis, PTCSpecNet model yielded a test set AUC of 0.991, cross-validation accuracy of 0.917, and recall of 0.979, while SVM achieved the highest test AUC (0.995) and LightGBM achieved the highest accuracy (0.978). For LNM prediction, LightGBM yielded the highest test AUC (0.916) and recall (0.909), whereas LymSpecNet model demonstrated competitive discriminative capacity (test set AUC = 0.913, cross-validation accuracy = 0.895) with an optimal clinical balance between precision (0.857) and recall (0.818). The fitting times for PTCSpecNet and LymSpecNet were 5.97 s and 8.52 s, respectively, while the estimated realistic clinical turnaround time is approximately 20–25 min, supporting rapid diagnostic feasibility. Conclusion: SERS analysis of FNA washout fluid combined with deep learning effectively captures the molecular fingerprint of the in situ thyroid nodule microenvironment, enabling high-sensitivity diagnosis of PTC and reliable prediction of LNM. Future prospective multi-center studies with external validation cohorts are warranted to further establish its broad clinical generalizability.
SN - 1176-9114
SN - 1178-2013
ER -