Fig. 11.

Evaluation of performance for AI in recognizing benign or malignant lesions. (a) The corresponding area under the ROC curve for the graphs is 90·6% for malignant pulmonary nodules versus benign nodules. The comparison between the predictive performance of model and human expert on another independent cohort of 284 patients was shown around the curve with corresponding false positive rate (TPR, sensitivity) and true positive rate (FPR, 1-specificity). (b) Contingency table for predicted labels and true labels of malignant and benign based on a cutoff score of our original model. (c-d) Ordered accuracy and error bar to assess the performance of AI when diagnosing the lesion status of lung nodules.