Skip to main content
. 2021 Jun;11(6):2354–2375. doi: 10.21037/qims-20-600

Table 6. Comparison of our multi-channel and multi-task deep learning (MMDL) method with the state of the art methods.

Cancer type Number of patients Images Extracted features AUC ACC (%) SEN (%) SPE (%)
A Coudray (1) NSCLC 567 Pathological Deep features 0.754
B Wang (26) LUAD 844 CT Deep features 0.810 73.86 72.27 75.41
C Xiong (27) LUAD 158 CT Deep features, clinical features 0.838 77.20 75.80 79.10
D Li (28) LUAD 1,010 CT Deep features, clinical features, radiomics 0.834 82.20 74.20
E Velazquez (10) LUAD 258 CT Radiomics, semantic features 0.670
F Liu (9) LUAD 288 CT Clinical features 0.709
G Li (21) NSCLC 312 CT Radiomics, clinical features 0.775
H Zhang (12) NSCLC 180 CT Radiomics 0.873 75.60 70.90 79.80
I Gevaert (20) NSCLC 186 CT Semantic features 0.890
J Guan (14) NSCLC 85 PET/CT SUVmax, clinical features 0.770 77.60 64.60 82.50
Our methods NSCLC 363 CT Deep features, clinical features 0.866 79.43 78.27 81.35

LUAD, lung adenocarcinoma; NSCLC, non-small cell lung cancer; EGFR, epidermal growth factor receptor; KRAS, Kirsten rat sarcoma; AUC, area under the curve; sensitivity; ACC, accuracy; SEN, sensitivity; SPE, specificity.