Fig. 1

Experimental workflow of multi-source feature model based on deep learning

Fig. 2

Segmentation images of MCN and SCN. (a) Abdominal axial T2WI showed high MCN signal with regular boundary. (b) Abdominal axial T2WI showed high SCN signal with irregular boundary

Fig. 2

The process of ResNet50 extracting deep learning features, which includes four convolution blocks (Conv), average pooling layer (Av-pool) and fully connected layer (fc)

Table 1

Single factor logistic regression analysis of clinical features of MCN and SCN

Table 2

Multi-factor logistic regression analysis of clinical features of MCN and SCN

Fig. 3

(a) Mean Square Error (MSE) graphs corresponding to different λ values of radiomics features. The MSE is minimized by adjusting the λ value to determine the optimal λ value; (b) Convergence diagram of the optimal characteristic coefficient of radiomics features; (c) Weight of radiomics features screened by LASSO, Skewness.1 represents skewness, Busyness.1 represents complexity, MCC.2 represents morphological correlation coefficient, DependencyVariance.2 represents dependency difference, Idn.4 represents inverse difference moment, and Correlation.7 represents correlation; (d) Mean square error (MSE) graphs corresponding to different λ values of deep learning features; (e) Convergence diagram of the optimal characteristic coefficient of deep learning features; (f) Weight of deep learning features screened by LASSO

Table 3

Performance of the feature models in the four classifiers

Fig. 4

ROC curve of four characteristic models. The abscissa is fpr (false positive rate) and the ordinate is tpr (true positive rate). The four classifiers are SVM, adaboost (ADAboost), randomforest (Random Forest) and logistic (Logistic). (a) ROC curve used by RAD feature model; (b) ROC curve of DL feature model; (c) ROC curve of RAD_DL feature model; (d) ROC curve of Clinical _RAD_DL feature model

Table 4

Perform a Delong test to compare the performance differences between the Clinical_RAD_DL feature model and other models and obtain the p-value

Fig. 5

Calibration curves and decision curves. (a) Calibration curves of RAD, DL, RAD_DL, and Clinical-RAD_DL feature model. The abscissa represents the PCN-predicted probability of PCN classification model, and the ordinate represents the observed probability; (b) Decision curves used by RAD, DL, RAD_DL, and Clinical-RAD_DL feature model. The abscissa represents the high risk threshold, and the ordinate represents the net benefit of the model, the All curve represents the net benefit predicted as MCN, and the None curve represents the net benefit predicted as SCN