Literature showed that cervical LNM negatively impacted patients’ overall survival and disease-free survival in PTC patients. It associated with higher rate of distant metastasis and disease-related mortality by 11.2-fold and 3-fold, respectively3, 5, 17. However, the detection rate of cervical LNM using US was unsatisfying, especially for central LNM. Besides, prophylactic cervical LN dissection could potentially lead to nerve injury and hypoparathyroidism17, 18. Therefore, preoperative prediction of cervical LNM was essential for all patients. Generally, US features like tumor size, echogenicity, “wider-than-taller” shape, extrathyroidal extension and calcification, clinical information like age were proved indicators of cervical LNM. However, because the expertise of the operator significantly influenced the diagnostic accuracy, these results only served as reference in clinical practice17-19.
With the development of radiomics analysis, several articles revealed a promising result for preoperative prediction of cervical LNM for PTC nodules. Liu introduced US radiomic analysis to preoperative cervical LNM prediction and proved that LNM was associated with larger size, younger age, irregular tumor shape, obscure boundary, spiculate margin, taller-than-wide shape, calcification, complex echo pattern, thyroid invasion and posterior region homogeneity (AUC in validation group = 0.782)19. In addition, Cui and Jiang reveal an AUC of 0.90 and 0.83 for cervical LNM prediction in the radiomics signature based on strain elastography ultrasound images and shear-wave elastography images. However, only Jiang’s study detected high-the predictive value of dimensional features and showed the wavelet transform of B-mode images were related with cervical LNM9, 18.
In the present study, with the help of ML, we established 10 models based on clinical characteristics and US radiomic features for preoperative prediction of cervical LNM in PTC nodules. The AUC and accuracy for each model varied in validation group and RF-RF model proved the best one among them. After cross-validation, RF-RF model selected the top 10 features of importance, which included age, 6 GLRLM, 1 GLCM, 1 shape feature and 1 first-order feature. Among them, 4 features were LOG based features, 3 were wavelet based feature and 1 were square root based feature. Unlike our study, Zhou et al demonstrated an ultrasound radiomics nomogram for central LNM with an AUC of 0.85 and their equation included age, TPOAB level, TG level, radiomic signature and ultrasonography-reported LN status. Because the radiomic features were extracted based on parameters mentioned by different guidelines, they lacked higher-dimensional features5. Besides, Tong et al established a nomogram for lateral LNM with an AUC of 0.91. They included 6 textural features (5 GLSZM and 1 GLCM) in the equation with little higher-dimensional features, either8.
Shape features described the shape of tumor volume, along with its geometric properties. For voxels within tumor volume, first-order features depicted their distribution of intensities while textural features measured their inter-relationship of distributions20. The GLRLM quantified the length of consecutive pixels of the same gray level value in images, while GLCM represented the number of times specific combination of gray levels occur in two separated pixels21. LOG filter acted as a combination of Laplacian operator and Gaussian filter, which might detect edges as well as noise in a smoothed image for filtering and differentiation22-24. It was employed for image filtering in the spatial domain and was widely used in radiomics in literature22, 24. Different filter sigma parameters applied for fine or coarse anatomic details for textural features22, 23. According to literature, wavelet filter could enhance certain characteristics based on its frequency domain in images and was widely used in image compression and preprocessing25, 26. Besides, square root filter could improve the overall condition of covariance matrices by improving their update accuracy and avoiding the negative definiteness27, 28. The LOG-, wavelet- and square root-filtered features selected by our result deeply implied the importance of including higher-dimensional statistical methods, meanwhile highlighting their unique role for radiomics analysis.
There were several limitations in our study. First, due to the retrospective nature, the clinical procedure was not strict and some nodules had incomplete clinical data. Besides, some of the recorded US images were discrete ones which might lead to the capture of unrepresentative portion of the tumor. Third, the number of samples was relatively small and larger cohorts were required for explorations in the future. Forth, models were validated in mono-center cohort. Thus, in order to have more convincing results, multi-center validation should be carried out in the future.
In conclusion, our study, for the first time, established preoperative prediction models with the help of ML for cervical LNM based on clinical characteristics and US radiomic features. They were expected to help with diagnosis, recurrence prediction and treatment decision of PTC.