Deep Learning-Assisted Diagnosis Of Focal Liver Lesions Based On Magnetic Resonance Imaging

Background and aims:Liver cancer is the sixth most common malignant tumor worldwide and the third leading cause of cancer-related deaths.In our country,the annual new cases and deaths of liver cancer account for more than 45% in the world.Early-stage diagnosis and treatment can improve survival rates of liver cancer patients.Early-stage diagnosis and treatment can improve survival rates of liver cancer patients.Dynamic contrast-enhanced MRI is the best imaging modality for liver tumor detection and differential diagnosis.However,in clinical practice,MRI diagnosis remains challenging,owing to liver tumor diversity and complex imaging features.The quality of diagnosis is uneven since it is affected by radiologists’ experience.Moreover,the high cost and long imaging acquisition time of full-protocol enhanced MRI inspection hampers its widespread use in screening and monitoring.Deep learning(DL),as an emerging artificial intelligence image recognition technology,may provide a new diagnostic strategy.In this study,convolutional neural networks(CNNs)were used to establish a deep learning system(DLS)to realize the assisted-diagnosis for focal liver lesions (FLLs)based on MRI images.The study has two parts: the first is to establish a deep learning-assisted diagnosis model based on MRI images and clinical information aiming to realize the automated accurate classification for FLLs,which was verified in an independent validation set and compared with experienced radiologists.The second is to explore the diagnostic value of DL models based on non-contrast sequences(T2,DWI)for FLLs,and further evaluated their performance on multi-center validation sets.Methods:In the first part of this study,we used data from 1,210 patients with liver tumors(N=31,608 images)to train CNNs to gain seven-way classifiers,binary classifiers,and three-way malignancy-classifiers(Model A-Model G).In three-way classifiers,we modified CNN structure to input clinical data.Models were validated in an external independent extended cohort of 201 patients(N=6,816 images).The area under receiver operating characteristic(ROC)curve(AUC)were compared across different models.We also compared the sensitivity and specificity of models with the performance of three experienced radiologists.In the second part of this study,a total of 50418 images of enhanced MRI from 1959 patients with liver tumor in three center were enrolled.T2-weighted imaging,diffusion-weighted imaging(DWI),and multiphasic T1-weighted imaging provided input for Inception-Res Net V2 through transfer learning to generate four models for three-way classification,i.e.benign lesion,primary Liver cancer and metastatic liver cancer.The models was then validated on the independent internal and two external datasets consisting of 5172 and 2916,1338 images,respectively.The diagnostic performance of non-contrast models(T2,T2+DWI)for distinguishing malignancy from benign tumors at-lesion/patient level were further evaluated in three validation sets.Results:In the first part of the study,deep learning achieves a performance on par with three experienced radiologists on classifying liver tumors in seven categories,i.e.cyst,hemangioma,focal nodular hyperplasia(FNH),other benign lesion,hepatocellular carcinoma(HCC),metastatic tumor,other primary liver cancer.Using only unenhanced images,CNN performs well in distinguishing malignant from benign liver tumors(AUC,0.946;95% CI 0.914–0.979 vs.0.951;0.919–0.982,P = 0.664).New CNN combining unenhanced images with clinical data greatly improved the performance of classifying malignancies as hepatocellular carcinoma(AUC,0.985;95% CI 0.960–1.000),metastatic tumors(0.998;0.989–1.000),and other primary malignancies(0.963;0.896–1.000),and the agreement with pathology was 91.9%.These models mined diagnostic information in unenhanced images and clinical data by deep-neural-network,which were different to previous methods that utilized enhanced images.The sensitivity and specificity of almost every category in these models reached the same high level compared to three experienced radiologists.Based on the findings of the first part,we further explored the diagnostic value of non-enhanced MRI sequences,and expanded the dataset coverage according to the clinical application scenarios.Non-contrast models demonstrated similar performance for classifying liver tumors to benign,primary malignant and metastatic tumors,compared with two models based on multi-sequence or enhanced images.In the independent internal cohort,the areas under the receiver operating characteristic curves of T2+DWI model reached 0.91(95% CI,0.888–0.932),0.873(95% CI,0.848-0.899),and 0.876(95% CI,0.840-0.911)for three categories,respectively.At patient-level,the sensitivities of malignant tumors gained from non-contrast models reached 98.1%,85.7%,87.5% respectively in three validation sets,while all specificities were almost greater than 70%.These results indicated that our models can identify over 95% patients with malignancy at best using non contrast images,while more than 70% of patients with benign tumors can avoid a further inspection using contrast mediums and the false negative rate is no more than 5%.Conclusions:1.DLS can accurately classify liver focal lesions into seven categories based on enhanced MRI images,and the addition of clinical information can significantly improve the precise classification performance of malignant tumors.DLS can be used as an assisted diagnostic tool for radiologists.2.Based on non-contrast MRI,deep learning algorithm can correctly classify liver tumors into benign,primary malignant and metastatic tumors at image-level,and distinguish malignancy from benign with high sensitivity at patient-level,potentially reducing scanning sequences and contrast agents to avoid side effects and financial costs,especially in patients without malignant lesions.With the assistance of DL,Non-contrast MRI is promising to used for screening,monitoring and follow-ups of potential liver focal lesions.