Application Of Deep Learning Based Artificial Intelligence In The Diagnosis Of Early Liver Cancer MRI

ObjectiveThe artificial intelligence(AI)algorithm based on the transfer attention mechanism and deep learning is used to post-process and analyze the characteristics of multiple sets of magnetic resonance images of patients with small liver cancer.The differences in image features between different sequences and their joint application are explored.An artificial intelligence multi-modal model for early diagnosis of liver cancer patients and its verification and machine learning.Materials and MethodsThe image of the patient group established by the artificial intelligence model is from 50 patients with early liver cancer nodules confirmed by surgery or ultrasound puncture pathology(the image is preoperative data).There are 61 nodules.The nodules all meet the clinical diagnosis criteria of early liver cancer.The model control group had no history of liver disease and no clinical manifestations.All the above patients and subjects used Siemens Skyra 3.0T and Philips Ingenia 3.0T superconducting MR magnetic resonance imaging images.The scans included conventional MRI and enhanced scan sequences and IVIM and QSM special sequences.The parameters of the T1WI fat-suppressed liver volume rapid acquisition sequence are:TR=3.1 ms.TE=1.5 ms,inversion time 5.0 ms.inversion angle =150°,matrix 320×256.layer thickness 5 mm,non-spacing scan,FOV=40 cm×32 cm.T2 fat saturation imaging parameters:TR=4064ms,TE=105ms,bandwidth 500Hz/point,inversion angle=150°,parallel image factor-2,scanning layers 28 layers,6mm each,matrix 256×256,pixels 1.3mm×1.3mm.After acquiring the images,the images of all patients were exported in DICOM format at the PACS workstation and numbered into a folder.Obtain the image and input itk-snap v3.8 software to segment the region of interest(ROI)to obtain binary images of each sequence.The output format is.NTTFI format.Algorithms and programs are written using Python software.,Compare feature extraction differences between sequences,and build an artificial intelligence image model.Patients were previously diagnosed with suspected intrahepatic nodules by imaging methods.They were divided into the machine group and the physician group to diagnose the nodules,and the results of the patients were followed up to obtain the sensitivity of the two groups of diagnostic results.Specificity,receiver operating characteristic curve(ROC curve)and AUC value were analyzed statistically using SPSS 26.0 software.The results were considered statistically significant at P<0.05.ResultsThe liver cancer model group and healthy model group feature extraction results were FS-T2WI(34.85 in the sHCC group and 15.33 in the healthy control group),DWI(20.36 in the sHCC group and 10.01 in the control group),and T1WI in the arterial phase(25.80 in the sHCC group,control 10.95 in the group)and T1WI in the venous phase(22.38 in the sHCC group and 11.96 in the control group).Differences in characteristics within and between groups were analyzed to meet normal distribution.The difference between T2WI in the sHCC group and DWI,T1WI in the arterial phase,and T1WI in the venous phase was statistically significant(p<0.001).There was difference between the three groups in DWI,T1WI in the arterial phase,and venous phase(p<0.01).There were significant differences between the four sequence groups(p?0.001)Nodules screening was performed on 100 patients with suspected intrahepatic nodules by two middle and senior doctors with 10-15 years of working experience,and the results were compared with the machine group.The result was the machine group false positive rate(3/71,0.04).False negative rate(1/42,0.02),sensitivity 93%,specificity 89%,physician group(5/71,0.07),false negative rate(3/42,0.07),sensitivity 87%,specificity 83%,The AUC values of the machine group and the physician group were 0.914 and 0.844,respectively.Conclusion1.Through the artificial intelligence model based on deep learning algorithm,effective feature extraction can be performed on intrahepatic nodule images of patients with small liver cancer,and the feature extraction results of FS-T2WI sequence have advantages.2.By comparing the intrahepatic nodules between the machine group and the physician group,the machine group has better accuracy in diagnosing small liver cancer,especially micro liver cancer(nodule maximum diameter ?1cm),and can perform effective machine learning.