| Objective:A group of soft tissue tumors（STTs）,which are easily misdiagnosed by imaging,will be used to establish an MRI texture analysis radiomics model for distinguishing benign and malignant by using their T1WI-FS、T2WI-FS、T1WI-FS+C images.The diagnostic efficiency of this model is compared with that of imaging specialists,aiming to explore the diagnostic value of the MRI texture analysis radiomics model based on the same MRI sequence in distinguishing benign and malignant STTs that are easilymisdiagnosed by imaging.Materials and Methods:1.Data CohortThe data of 65 consecutive patients with soft tissue tumors whose MRI findings met the definition of STTs which were easily misdiagnosed on imaging were collected from the Picture Archiving and Communication System（PACS）of the Second Hospital of Da Lian Medical University from March 2017 to October 2020,including 34 men and31 women,with an mean age of 51.31±19.62 years（age range,5-85）.2.MRI equipment and examination methodSixty-five patients were scanned with GE Discovery MR750W 3.0T and Siemens Magnetom Verio 3.0T magnetic resonance scanner to obtain T1WI-FS、T2WI-FS、T1WI-FS+C sequences,and the coils were adjusted according to the lesion site.3.MRI imaging indicators and the results of imaging specialistsTwo imaging specialists used a double-blind method to analyze all images,recorded the corresponding imaging observation indicators,and classified these 65patients into benignity and malignancy.The imaging observation indicators included:tumor size,depth,signal,boundary,peritumoral changes,presence or absence of（false）capsule,adjacent bones or blood vessels and nerves were involved or not.If thediagnosis results of two specialists were inconsistent,a third senior specialist joined the joint discussion and reached a consistent result.4.The construction of texture analysis radiomics modelImporting the original DICOM images（T1WI-FS、T2WI-FS、T1WI-FS+C）of 65soft tissue tumors into the radiomics cloud platform,and manually delineated the area of interest layer by layer along the edge of the lesion on the above three sequences,the computer automatically generated 3D volume of interest（VOI）and obtained texture features.The optimal texture features which selected by variance selection,univariate feature analysis and selection,and LASSO algorithm were used to build SVM（Support Vector Machine）and LR（Logistic Regression）machine learning classifier models.5.Statistical AnalysisSPSS 26.0 was used for statistical analysis,categorical variables were represented by frequency,Fisher’s exact test or chi-square test was used at first,and then binary logistic regression method was used.The diagnostic efficacy evaluation of the imaging specialist was achieved by drawing the ROC curve and calculating the area under curve（AUC）,sensitivity and specificity.The effectiveness evaluation of the MRI texture analysis radiomics model obtained the ROC curve of SVM and LR through the radiomics platform and calculated the AUC value,accuracy,sensitivity,and specificity,and calculated the accuracy,recall rate and f1-score at the same time.The pros and cons of the diagnostic efficiency between the MRI texture analysis radiomics model and the imaging specialists were evaluated by the Delong test.Results:1.MRI manifestations of STTs easily misdiagnosed on imagingThe binary logistic regression showed that the factors related to the misdiagnosis of benign and malignant STTs were the depth of the tumor,whether the tumor signal was uniform and whether the boundary was clear（P value<0.05）.2.Diagnostic performance of imaging specialistsThe diagnostic accuracy of imaging specialists in identifying benign and malignant STTs easily misdiagnosed by imaging is 0.52,AUC is 0.52,sensitivity is 0.56,and specificity is 0.52.3.Diagnostic performance of MRI texture analysis radiomics modelThe T1WI-FS、T2WI-FS、T1WI-FS+C sequences of 65 patients with STTs were performed texture analysis respectively,SVM and LR classifiers were constructed using the optimal texture feature values obtained by dimensionality reduction.The performance of the SVM machine learning classifier was slightly better than that of the LR classifier.The results of Delong test showed that there was no significant difference in AUC between the four machine learning classifier models of SVM（P>0.05）.4.The diagnostic performance comparison between imaging specialists and radiomics modelAfter Delong test,the difference in ROC between the imaging specialists and the SVM machine learning classifier based on the T1WI-FS、T2WI-FS、T1WI-FS+C combined sequence was statistically significant（P=0.081）.The diagnostic efficiency of the latter（AUC=0.85）was significantly better than the former（AUC=0.52）.Conclusion:Through the texture analysis of T1WI-FS、T2WI-FS、T1WI-FS+C sequences MR images of 65 cases of STTs that were easily misdiagnosed by imaging,the MRI texture analysis radiomics model was constructed,and its diagnostic efficiency was compared with the diagnosis of imaging specialists,the following conclusions can be drawn:1.The radiomics model constructed based on the MRI texture analysis of the three sequences of T1WI-FS、T2WI-FS、T1WI-FS+C can be used to distinguish benign and malignant soft tissue tumors that were easily misdiagnosed by imaging.2.Compared with the diagnostic performance of the imaging specialists,the MRI texture analysis radiomics model based on the same sequence has a higher differential diagnosis performance.The MRI texture analysis radiomics model has a diagnostic power（AUC）of 0.85,and the diagnostic performance of the imaging specialists（AUC）is 0.52. |
