Cervical LNM is very common for PTC patients [3]. The inaccurate preoperative diagnosis of metastatic LNs can lead to insufficient surgical scope, potentially leaving metastatic LNs behind and increasing the risk of recurrence. Moreover, it may also result in an overly extensive surgical scope, increasing the risk of surgical complications, particularly in the case of lateral cervical LNM [3, 4]. Because the more extensive dissection required for lateral cervical LNM is associated with cosmetic concerns, such as a longer incision, and the potential for nerve injury (accessory, marginal mandibular, sympathetic nerves, et al.), hemorrhage, and chyle leak, which are not associated with central cervical LN dissection [4]. Therefore, an accurate preoperative diagnosis of lateral cervical LNM contributes to defining the appropriate surgical scope and reducing the recurrence rate [34, 35]. Although US is the primary imaging technique in the diagnosis of thyroid nodules and cervical LNM, its diagnostic efficacy is limited [5, 9, 10]. Therefore, it is crucial to develop an effective technique to improve the diagnostic efficacy of US for lateral cervical LNM.
Previous studies have developed two types of models to improve the diagnosis of lateral cervical LNM. The first type is predictive models that indirectly predict LNM with radiomics features such as the size, shape, boundary, and calcification of LNs on US images [36, 37]. Park et al. analyzed US images of primary thyroid tumors and obtained 14 radiomics features, constructing a radiomics-based model with an AUC of 62.1% in predicting lateral LNM [36]. Similarly, Liu et al. developed a predictive model for cervical LNM by selecting 50 radiomics features on US images, which showed an AUC of 72.7% and an accuracy of 71.0% in an independent testing cohort [37]. Overall, these radiomics-based predictive models showed limited improvement in diagnostic efficacy. Moreover, while predictive models can assist in determining the surgical method for patients, such as whether to perform lateral cervical LN dissection, they cannot pinpoint the specific level within level I to V where metastatic lymph nodes may be located. Therefore, their contribution to guiding the surgical scope for patients with lateral LNM is limited. In contrast, the second type of model, the diagnostic model, can avoid this problem, which directly diagnoses lateral cervical LNM by analyzing the US images of LNs. It is based on a deep learning algorithm that has been widely applied in medical image recognition and has shown promising diagnostic efficacy [38, 39]. Lee et al. retrospectively analyzed the LN US images from 804 patients with PTC, developing a CAD system using a CNN Cmodel that classified lateral LNs with promising performance [19]. Although their model can infer the locations of LNs by drawing attention heatmaps, the important regions need to be cropped from the original US images, which is a labor- and time-intensive task. Overall, it failed to automatically detect LNs on the original US images.
The CNN model was introduced by LeCun Y et al. in 1998 and has been widely used in medical image diagnosis since then [40]. With the development of AI, some excellent deep learning models have been designed. Among those models, the Transformer model, introduced by Vaswani A et al. in 2017, possesses a unique self-attention mechanism that facilitates the capture of long-range dependencies within input images [41]. Our results revealed that compared to other traditional CNN models, the Deformable DETR model showed the best performance in the detection and classification of LNs. This is attributed to its Transformer-based architecture, which allows the accurate identification of smaller targets within a larger context [30]. Regarding the performance in LN detection, the CAD system was comparable to the senior sonographer and significantly superior to the intermediate sonographer and junior sonographer in the retrospective and prospective test sets. However, similar to the three sonographers, the CAD system also missed some LNs. Interestingly, we found that Level V lymph nodes were the most likely to be missed. The reason may be that the cervical Level V lymph nodes are located at the posterior side of the neck, bordered by the posterior edge of the sternocleidomastoid muscle, the anterior edge of the trapezius muscle, and the upper edge of the clavicle [42, 43]. They are situated deeper and are relatively smaller in size. In a horizontal comparison of the accuracy of the model and the three sonographers in detecting lymph nodes in Level V, it was found that the model had stronger locating ability for Level V lymph nodes compared to the junior and intermediate sonographers, but due to the small sample size, this difference was not statistically significant. Overall, sonographers should use the detection results of the CAD system to further improve their performance in detecting LNs. Regarding the LN classification, the CAD system also showed excellent classification efficacy, with an AUC of 94.4% in the retrospective test set, greatly outperforming all sonographers. Lee et al. also developed a CAD system based on a traditional CNN model to classify lateral cervical LNs, achieving an accuracy, sensitivity, and specificity of 83.0%, 79.5%, and 87.5%, respectively [19]. Compared to their model, the CAD system based on the Deformable DETR model proposed here showed significantly better classification performance. Additionally, the CAD system can also automatically detect lateral cervical LNs on original US images. Notably, evaluating the diagnostic capabilities of the CAD system and sonographers should involve both detection and classification performance for cervical LNs. For instance, the intermediate sonographer showed a lower AUC compared to junior sonographer in the prospective test set, which did not mean the junior sonographer possessed superior diagnostic capabilities. This may attribute to the fact that junior sonographer detected fewer LNs. The CAD system had reliable diagnosis capabilities because of its superior performance in both detection and classification, compared to the sonographers. Furthermore, we observed that the CAD system and sonographers had better detection and classification performance in the prospective test set compared to the retrospective test set. This discrepancy may be attributed to the relatively limited data available in the prospective test set. In the future, we intend to recruit a larger dataset to comprehensively validate the diagnostic efficacy of the CAD system.
Despite the achievements of the proposed CAD system, the diagnostic efficacy remains to be consistently improved in future studies by conducting a large-scale multicenter study and incorporating additional LN data across various brands of US instruments. Moreover, as all included US images were static in this study, the clinical application of the CAD system in dynamic US is unknown but may holds substantial clinical value.