Application Of Deep Learning-based Ultrasound Radiomics For The Characterization And Source Prediction Of Lymphadenopathy

Objective: The referral patterns and treatment strategies of lymphadenopathy caused by different etiologies such as reactive hyperplasia,tuberculous lymphadenitis,lymphoma and metastatic lymph nodes from malignant tumors are very different.Timely and accurate preoperative diagnosis of the specific etiology of lymphadenopathy and prediction of the primary tumor origin of metastatic lymph nodes can reduce the delay in diagnosis and unnecessary invasive examination.However,current ultrasound qualitative lymphadenopathy is usually based on a small number of subjective visual features of a single ultrasound modality,which limits sonographer diagnostic accuracy.In this study,cervical lymphadenopathy(CLA)was selected as the research area to explore the application value of multi-modal ultrasound intelligent radiomics based on deep learning(DL)for characterization and source prediction of lymphadenopathy.Based on B-mode ultrasound(BUS)and Color Doppler Flow Imaging(CDFI)images,a unexplained CLA hierarchical diagnostic model(CLA-HDM)was constructed.The model uses DL algorithms to extract high-throughput quantitative features from bimodal ultrasound images to classify four common etiologies of CLA(reactive hyperplasia,tuberculous lymphitis,lymphoma and metastatic CLA)and to assess the value of the model as a clinical diagnostic aid for ultrasonographers of different levels of experience.Subsequently,a DL multiple step modality fusion model(MSMFM)for the diagnosis of squamous cell carcinoma(SCC)and adenocarcinoma(ADC)subtypes of metastatic CLA was constructed based on BUS images,CDFI images,ultrasound elastography(UE)images,contrast-enhanced ultrasound(CEUS)videos and patient clinical information of metastatic CLA to provide a preliminary basis for noninvasive preoperative prediction of the primary tumor origin of metastatic CLA.Finally,MSMFM was further trained to construct an intelligent model for predicting the origin of metastatic CLA primary tumors and to evaluate its performance in predicting the primary tumors as head and neck squamous cell carcinoma(HNSCC),thyroid cancer(TC),lung cancer(LC),and gastrointestinal cancer(GIC).Methods: 1.BUS images,CDFI images and pathological diagnosis results of763 patients with unexplained CLA who underwent ultrasound-guided cervical lymph node biopsy in three hospitals were retrospectively collected and divided into training set(n=395),internal testing set(n=171),external testing set 1(n=105)and external testing set 2(n=92)according to the collection hospital and collection time order.By simulating the sonographer ’s clinical diagnostic thinking,three DL sub-models with the deep convolutional neural network(DCNN)as the initial framework were first constructed based on the BUS images and CDFI images of patients in the training set;among them,sub-model 1 performed primary diagnosis to distinguish benign from malignant CLA,sub-models 2 and 3 performed secondary diagnosis to distinguish reactive hyperplasia and tuberculous lymphadenitis in benign CLA and lymphoma and metastatic CLA in malignant CLA,respectively;the latter three sub-models were integrated into CLA-HDM to directly identify the specific etiology of each unexplained CLA.The sub-models and the integrated CLA-HDM were validated on both internal and external testing sets.Six different experienced sonographers(two senior,two middle,and two junior)diagnosed patients on the same testing sets before and after CLA-HDM assistance,and the clinical assistance diagnostic value of the integrated intelligent model was assessed by analyzing changes in sonographers performance.2.The BUS images,CDFI images,UE images,CEUS videos and clinical baseline data of 301 patients with metastatic CLA(121 SCC,180 ADC)who underwent ultrasound-guided cervical lymph node biopsy and subsequently had a definite pathological subtypes at the Second Hospital of Lanzhou University from November 2018 to June 2021 were retrospectively collected.Univariate analysis of clinical baseline data from patients with different metastatic CLA pathological subtypes was performed to screen for key clinical indicators.Based on the imaging characteristics of the four ultrasound modalities and the clinical experience of the sonographers,DL algorithm was used to extract the heterogeneity features among the modalities to build an intelligent MSMFM fusing static images(BUS images,CDFI images,UE images),dynamic CEUS videos and key clinical indicators of the patients.Three-fold cross-validation method was used to train and evaluate the model,and the performance of MSMFM,different ultrasound modalities,different ultrasound modality combinations,and two senior sonographers in the diagnosis of SCC and ADC subtypes in metastatic CLA patients was compared.3.The BUS images,CDFI images,UE images,CEUS videos,and clinical baseline data were retrospectively collected from 280 patients with metastatic CLA(54 primary tumors with HNSCC,58 primary tumors with TC,92 primary tumors with LC,76 primary tumors with GIC)who underwent ultrasound-guided cervical lymph node biopsy and subsequently had a definite primary tumor origin at the Second Hospital of Lanzhou University from February 2019 to October 2021,and divided into training set(n=185)and testing set(n=95)according to the chronological order of collection.Univariate analysis of clinical baseline data from patients with metastatic CLA of different primary tumor origin in the training set was performed to establish the clinical model.Based on conventional ultrasound(CUS;including BUS and CDFI)image data,the performance differences of DL models based on CUS,CUS + UE,CUS + CEUS and CUS + UE + CEUS data in predicting primary tumors as HNSCC,TC,LC and GIC were compared to select the best modal combination and fused with clinical models to construct a comprehensive prediction model.The comprehensive predictive model was validated on the testing set and the performance of the model in predicting each specific primary tumor origin from metastatic CLA ultrasound images was evaluated.Results: 1.All three task-specific sub-models and the CLA-HDM showed good performance in classifying patients with unexplained CLA.In the training set,internal testing set and external testing set 1 and 2: the area under the curve(AUC)for sub-model 1 to distinguish benign and malignant unexplained CLA was 0.986,0.932,0.963 and 0.896,respectively;the AUC for sub-model 2 to distinguish tuberculous lymphadenitis from reactive hyperplasia was 0.935,0.922,0.857,and 0.872,respectively;the AUC for sub-model 3 to distinguish lymphoma from metastatic CLA was 0.979,0.852,0.847,and 0.827,respectively;and the AUC for integrated CLA-HDM to distinguish reactive hyperplasia,tuberculous lymphadenitis,lymphoma,and metastatic CLA was 0.964,0.873,0.837,and 0.840,respectively.CLA-HDM had an AUC of 0.718,0.875,and 0.812 for the diagnosis of reactive hyperplasia;0.883,0.860,and 0.897 for the diagnosis of tuberculous lymphadenitis;0.816,0.670,and0.936 for the diagnosis of lymphoma;and 0.855,0.825,and 0.804 for the diagnosis of metastatic CLA in the three testing sets.CLA-HDM was more accurate overall than the six sonographers with different levels of experience.The accuracy,sensitivity,and specificity of these six sonographers generally improved with CLA-HDM assistance,especially for junior and middle-aged sonographers,who improved performance after model assistance comparable to that of senior sonographers before model assistance(P >0.05),and assisted in reducing the diagnostic false positive rate by 0.7-3.1% and the diagnostic false negative rate by 2.2-10%.2.Compared with single,two or three-modality ultrasound fusion models,four-modality ultrasound fusion model(BUS + CDFI + UE + CEUS)performed better in the diagnosis of SCC and ADC subtypes in metastatic CLA patients(single-modal vs.two or three-modal vs.four-modal = 0.618～0.761 vs.0.667～0.792 vs.0.838).Univariate analysis of patient gender(P < 0.001),age(P = 0.005),lymph node location(P = 0.010),neck partition(P < 0.001),longitudinal diameter(P < 0.001)and transverse diameter(P = 0.019)were screened as key clinical indicators.When the key clinical indicators were superimposed on the four ultrasound modality fusions,the constructed MSMFM showed the best diagnostic performance(AUC of 0.857 and accuracy of 80.1%)and was significantly better than the level of two senior sonographers(accuracy of45.9%～46.8%,P < 0.05).3.Univariate analysis showed that age(P < 0.001),sex(P =0.008),and lymph node neck partition(P < 0.001)were statistically different between patients with metastatic CLA of different primary tumor origin and were used to establish the clinical model.Compared with clinical,CUS,CUS + UE,CUS + CEUS and CUS + UE + CEUS models,the comprehensive prediction model had the best performance in predicting the primary tumor site of metastatic CLA(clinical vs CUS vs CUS + UE vs CUS + CEUS vs CUS + UE + CEUS vs comprehensive prediction =0.630 vs 0.708 vs 0.721 vs 0.741 vs 0.755 vs 0.822).The AUC and accuracy of this combined model for predicting metastatic CLA from HNSCC,TC,LC,and GIC sites were 0.869 and 85.1%;0.838 and 87.3%;0.750 and 67.3%;and 0.829 and 77.8%,respectively.Conclusion: 1.The DL intelligent CLA-HDM proposed in this study based on BUS images and CDFI images has high accuracy,sensitivity and specificity in the classification and diagnosis of four common etiologies of unexplained CLA,and can assist in shortening the diagnostic gap between sonographers with different levels of experience and reducing unnecessary diagnostic delays and invasive examinations.2.The DL Intelligent MSMFM constructed by combining the advantages of four ultrasound modalities and integrating clinical information has a significantly higher performance than experienced sonographers in the diagnosis of SCC and ADC subtypes in patients with metastatic CLA.It is expected to assist in accelerating the localization of the primary tumor in clinical practice and provide active guidance for subsequent treatment.3.The DL comprehensive prediction model established based on multimodality ultrasound data and clinical information has a good predictive performance for the source of the primary tumor in patients with metastatic CLA and can be a potential tool to guide the personalized diagnosis and treatment of metastatic CLA patients.