| Objective: The difference of CT morphological signs,quantitative parameters and prediction scores among lung adenocarcinoma in different pathological classification was evaluated by the artificial intelligent assisted diagnosis system based on deep learning,to analyze the application value in pathological classification of the lung adenocarcinoma with the diameter less than or equal to 3 cm and evaluate its predictive performance.Methods: CT images of 713 patients with lung adenocarcinoma within 3 cm in diameter which were confirmed by pathology were retrospectively analyzed,Then these cases were randomly classified as training group and verification group in a ratio of two to one,there were 476 cases in the training group and 237 cases in the verification group respectively.Based on pathological type,the verification group data were divided into three groups:preinvasive lesions(AAH and AIS),microinvasive adenocarcinoma(MIA)and invasive adenocarcinoma(IAC).Patients were scanned on Discovery CT750 HD CT scanner using a low-dose spiral chest scanning mode.Scanned images were reconstructed at 1.25 mm slice thickness and uploaded to the PACS system.FCN(Fully Convolutional Networks),end-to-end target detection model(Single Shot Multibox Detector,SSD)and LSTM(Long Short-Term Memory)blend together in the two-stage cascade 3D-CNN model.CT morphological features of 476 cases in the training group were identified by three senior radiologists,then these cases were divided into four grades depending on the infiltration degree of lung adenocarcinoma.A deep-learning model was trained,and the 237 cases of verification group data were used to test it.The morphological signs of the lesions in the verification group were analyzed,and the CT quantitative parameters were measured and calculated automatically including three-dimensional maximum length diameter,short diameter,volume and maximum CT value,minimum CT value,mean CT value.The pathological types of lung adenocarcinoma in the verification group were graded on a scale of 2-5.The differences of parameters among the groups were compared.The differences among three groups were compared by LSD-t method,the statistical data were analyzed by chi-square test or Fisher’s exact probability method.Draw ROC curve and calculate the area under the curve,analyze the diagnostic performance of each indicator and establish the optimal diagnostic threshold.Results: In the CT quantitative parameters among three different pathological classification of lung adenocarcinoma,differences of the three-dimensional maximum major axis,minor axis,volume,maximum CT value,average CT value and predicted score calculated by the artificial intelligence among the three groups were statistically significant(P < 0.05).The difference of the minimum CT value was not statistically significant(P>0.05).In the morphological signs among three different pathological classification of lung adenocarcinoma,there were statistically significant differences in lobulation,spiculation,vascular convergence and pleural tag diagnosed by artificial intelligence among the three groups(P < 0.05),but there was no significant difference in vacuolar sign(P > 0.05).Score > 3.5 predicted by artificial intelligence were considered as the critical point for the diagnosis of invasive lesion,the area under the curve(AUC)was 0.873,the diagnostic sensitivity was 81.9% and specificity was 84.4%.In addition,the AUC values of three-dimensional maximum length diameter,short diameter,volume,maximum CT value and average CT value were0.873,0.890,0.879,0.892,0.805,0.719,these parameters have great diagnostic performance.Conclusions: The artificial intelligence auxiliary diagnosis system based on deep learning has certain predictive value in differentiating pathological type of lung adenocarcinoma less than or equal to 3 cm in diameter.Additionally,it has good repeatability.Meanwhile,it can effectively improve the accuracy of pathological classification of lung adenocarcinoma. |
PDF Full Text Request