Deep Learning-based Contrast Enhanced Ultrasound In Diagnosing Focal Liver Lesions: A Prospective,Multicenter Study

Purpose:To assess the diagnostic efficacy of standardized contrast enhanced ultrasound(CEUS)based on on-site image reading in a daily practice scenario prospectively collected from multi-centers across China,so as to provide high-quality evidence.To develop and validate deep learningbased CEUS models in differentiating hepatocellular carcinoma(HCC)and intrahepatic cholangiocellular carcinoma(ICC),and evaluate whether assistance from artificial intelligence(AI)can lead to an improvement for radiologists.Before the construction of the deep learningbased models,we committed to develop a visual-tracking framework for CEUS data using modified fully-convolutional siame(Siam FC)method to solve disadvantages of relying on laboriously and inefficiently manual segmentation.Materials and Methods:⑴ A total of 80 centers distributed in 24 provinces or municipalities(24/34,71% of total provinces and municipalities in China)across the country uploaded CEUS data to the online database,including 3876 patients with 3960 lesions.The efficacy of CUES was evaluated via on-site reports,and the sensitivity(Se),specificity(Sp),positive predictive value(PPV),negative predictive value(NPV)and accuracy(ACC)were calculated.Subgroup analysis was performed according to lesion size,liver background,hospital rank and radiologist experience to explore the potential factors affecting diagnostic efficacy.⑵ We made two improvements,including a video attention methodology with an object location-point restriction and innovative improvements to the traditional linear template updating,to the Siam FC framework to elevate its robustness to adapt to our region of interest(ROI)tracking task.Our automatic tracking of models was constructed based on 875 cases of CEUS data prospectively collected from multi-centers with precise labels,and the performances of accuracy,robustness and speed aspects were compared to traditional methods.⑶ The deep learning-based CEUS models were constructed to discriminate HCC and ICC,and the models were developed successively from gray imagines,CEUS imagines,and combination of gray and CEUS,and finally integrated clinical factors to diagnose FLLs.A total of 1005 patients were randomly divided into training and validation cohorts with a 3:1 ratio.Two independent cohorts with 97 and 26 patients from two hospital outside the multi-centers were enrolled as test cohorts.The accuracy and robustness of model were investigated by validation and test cohorts.Finally,we compared the model’s performance to radiologists with different levels,and further compared the diagnostic efficacy of radiologists with or without assistance from AI.Results:⑴ For differentiating malignant from benign FLLs,the Se,Sp,PPV,NPV and ACC were96%,84.9%,94.5%,88.7% and 93.0%,respectively,and the good performance continued in subgroups analysis.The specificity and accuracy was superior in lesions >5 cm group when compared to groups of lesions ≤2 cm and 2-5 cm.Compared to the normal group,specificity in hepatic fibrosis/cirrhosis group decreased(87.9% Vs 72.5%),while sensitivity increased(94.0%Vs 97.2%),and the accuracy increased(91.5% Vs 94.8%)(p<0.001).When compared to non-expert centers,sensitivity in expert centers increased while specificity decreased(p<0.001),and accuracy was slightly higher in expert centers than non-expert centers(93.6% Vs92.0%)(p=0.028),while no difference was found among radiologists with different levels.For differentiating five common FLLs,including HCC,metastasis,abscess,hemangioma and focal nodular hyperplasia(FNH),the sensitivities were 89.4%,85.2%,94.5%,82.8% and 91.3%,respectively,however,sensitivity for ICC was only 54.3%-56.3%,and up to 18% ICC(42/232)was misdiagnosed to HCC.⑵ We constructed a Siam FC framework to adapt to our region of interest(ROI)tracking task,and the performance of the automatic tracking models was validated in 31 CEUS videos with precisely labeled CEUS data(4367 frames),and the performance of accuracy,robustness and speed aspects was compared to traditional methods.The results showed that the average ACC reached to 85.7%,the robustness was increased significantly at all thresholds of intersection over unio(IOU),which increased up to 25% at the value IOU equaled to 0.1,and the speed reached to 99 frames per second,which would be definitely competent for real-time dynamic tracking.⑶ In the validation cohort,to identify HCC from ICC,AI achieved an area under the curve(AUC)of 0.889(95% CI 0.890–0.978),and the Se,Sp,ACC were 0.914(95%CI:0.912-0.916),0.804(95%CI:0.798-0.810)and 0.890(95%CI:0.888-0.892),respectively.When compared to radiologists reviewing CEUS videos along with complementary patient information,AI outperformed all levels of radiologists(ACC 0.813;95%CI:0.765-0.856)(P <0.05).With the assistance of AI,radiologists exhibited a significant improvement in all levels(ACC 0.880;95%CI: 0.838-0.915)when compared to scenarios without AI assistance(ACC0.813;95%CI: 0.765-0.856)(P < 0.001),and the improvement was the greatest in junior radiologists.Conclusions:⑴ CEUS shows a favorable on-site diagnostic ability for differentiating malignant and benign FLLs,and the diagnostic efficacy was almost unaffected by lesion size,liver background,hospital rank and radiologist experience.It obtains satisfactory sensitivity for HCC,metastasis,abscess,hemangioma and FNH,while with a low sensitivity for ICC.⑵ The automatic tracking models with Siam FC framework obtained good performance of accuracy,robustness and speed aspects when compared to traditional methods,which would be definitely competent for real-time dynamic tracking.⑶ Deep learning based CEUS models can achieved a favorable AUC,Se,Sp and ACC,and AI outperformed all levels of radiologists.With the assistance of AI,radiologists exhibited a significant improvement in all levels,and the improvement was greatest in junior radiologists.