The diagnostic and prognostic value of radiomics and deep learning technologies for patients with solid pulmonary nodules in chest CT images

Table 2 Predictive performance of different classification models and junior radiologists in the testing set

Model or radiologist	Sensitivity	Specificity	Accuracy	AUC
RF + Clinical	0.535 [0.437, 0.633]	0.740 [0.656, 0.824]	0.640 [0.574, 0.706]	0.721 [0.651, 0.791]*
RF + Radiomics	0.747 [0.661, 0.833]	0.606 [0.512, 0.700]	0.675 [0.611, 0.739]	0.778 [0.738, 0.858]
RF + Combined	0.616 [0.520, 0.712]	0.788 [0.709, 0.867]	0.704 [0.641, 0.767]	0.811 [0.713, 0.839]
CNN	0.758 [0.674, 0.842]	0.788 [0.709, 0.867]	0.773 [0.715, 0.831]	0.816 [0.758, 0.875]
CNN+ Clinical	0.778 [0.696, 0.860]	0.788 [0.709, 0.867]	0.783 [0.726, 0.840]	0.819 [0.760, 0.877]
Radiologist 1	0.778 [0.696, 0.860]	0.452 [0.356, 0.548]	0.611 [0.544, 0.678]	0.615 [0.538,0.692]
Radiologist 2	0.990 [0.970, 1.000]	0.519 [0.423, 0.615]	0.749 [0.689, 0.808]	0.755 [0.688,0.821]

*Significant difference was found between the CNN model with clinical features and RF with clinical features by Delong test (p < 0.05)
Abbreviations: RF Random forest, CNN Convolutional neural network, AUC Area under the receiver operating characteristic curves

ISSN: 1471-2407