TABLE III. Ablation Studies on the Feature Extractors in 3-Label (Normal vs. Luad vs. Lscc) Classification Task.
We Used Different Feature Extractors for Graph Construction and Evaluated Their Role on the Overall Classification Task. Here, Resnet* Indicates the Use of a Pre-Trained Resnet18 Network Without Fine-Tuning, Resnet† Indicates With Fine-Tuning. Cae is a Convolutional Auto Encoder. CL Represents Contrastive Learning on STL10 or Nlst by Our Method. Mean Performance Metrics Are Reported Along with the Corresponding Values of Standard Deviation in Parentheses
| (a) Precision, Recall/Sensitivity, and Specificity (Percentage (%) values are reported). | ||||||||||
|
| ||||||||||
| Method | Data | Precision | Recall/Sensitivity | Specificity | ||||||
| Normal | LUAD | LSCC | Normal | LUAD | LSCC | Normal | LUAD | LSCC | ||
|
| ||||||||||
| Resnet* | CPTAC | 88.3(4.2) | 63.8(2.7) | 76.4(5.0) | 81.1(5.9) | 71.0(4.1) | 73.6(7.6) | 94.0(3.1) | 80.9(1.8) | 88.4(3.6) |
| TCGA | 56.2(7.2) | 42.6(3.1) | 35.6(10.3) | 50.9(11.3) | 31.3(24.4) | 50.1(22.8) | 79.7(8.6) | 77.7(19.8) | 58.3(11.3) | |
|
| ||||||||||
| Resnett | CPTAC | 88.1(5.1) | 77.6(5.6) | 77.7(4.1) | 89.4(3.5) | 72.2(6.6) | 80.1(6.6) | 93.3(3.6) | 89.6(3.8) | 88.3(3.6) |
| TCGA | 62.1(2.1) | 51.8(3.6) | 66.6(5.6) | 77.8(3.2) | 49.5(9.3) | 52.5(6.4) | 77.6(2.9) | 76.3(5.9) | 85.6(4.8) | |
|
| ||||||||||
| CAE | CPTAC | 90.6(1.7) | 77.8(2.2) | 73.2(2.7) | 88.5(4.2) | 65.6(4.5) | 85.8(3.6) | 95.1(1.0) | 91.1(1.2) | 84.3(2.9) |
| TCGA | 61.6(2.3) | 42.3(2.5) | 57.2(2.1) | 78.3(3.2) | 35.9(1.6) | 51.1(5.7) | 76.9(2.7) | 75.1(2.1) | 80.0(1.8) | |
|
| ||||||||||
| CL (STL10) | CPTAC | 95.3(1.6) | 87.6(3.5) | 91.7(3.4) | 96.2(1.6) | 90.5(4.2) | 87.3(3.7) | 97.5(0.9) | 93.8(2.2) | 95.9(1.8) |
| TCGA | 76.9(2.7) | 66.5(3.3) | 73.8(3.5) | 82.2(5.5) | 73.4(1.5) | 61.0(4.1) | 88.4(1.8) | 81.0(3.1) | 88.5(2.8) | |
|
| ||||||||||
| CL (NLST) | CPTAC | 93.2(3.0) | 88.4(3.9) | 87.8(3.0) | 95.9(2.2) | 83.9(4.5) | 89.2(4.0) | 96.2(1.7) | 94.7(1.9) | 93.8(1.7) |
| TCGA | 89.2(2.8) | 74.4(2.7) | 84.4(0.7) | 92.6(2.7) | 79.8(1.9) | 75.2(1.6) | 94.7(1.6) | 86.0(2.3) | 92.7(0.4) | |
|
| ||||||||||
| (b) Accuracy and AUC (Percentage (%) values are reported). | ||||||||||
|
| ||||||||||
| Method | Data | Accuracy | AUC | |||||||
|
| ||||||||||
| Resnet* | CPTAC | 75.4(2.0) | 88.9(0.4) | |||||||
| TCGA | 44.0(2.8) | 60.4(3.0) | ||||||||
|
| ||||||||||
| Resnett | CPTAC | 80.8(1.1) | 92.2(0.9) | |||||||
| TCGA | 59.6(1.1) | 76.2(0.8) | ||||||||
|
| ||||||||||
| CAE | CPTAC | 80.2(1.6) | 93.2(0.9) | |||||||
| TCGA | 54.7(2.2) | 72.9(1.6) | ||||||||
|
| ||||||||||
| CL (STL10) | CPTAC | 91.4(1.1) | 98.0(0.9) | |||||||
| TCGA | 71.9(1.9) | 86.4(1.6) | ||||||||
|
| ||||||||||
| CL (NLST) | CPTAC | 91.2(2.5) | 97.7(0.9) | |||||||
| TCGA | 82.3(1.0) | 92.8(0.3) | ||||||||
|
| ||||||||||
| (c) DeLong’s test to compare the AUCs of models generated using different feature extractors. log10(0.05) = −1.301. | ||||||||||
|
| ||||||||||
| Method | Data | log10(P-value) | ||||||||
|
| ||||||||||
| Resnet* | CPTAC | −3.711(2.044) | ||||||||
| CPTAC | −3.711(2.044) | |||||||||
|
| ||||||||||
| Resnett | CPTAC | −2.336(0.592) | ||||||||
| TCGA | −4.041(0.418) | |||||||||
|
| ||||||||||
| CAE | CPTAC | −2.325(2.573) | ||||||||
| TCGA | −4.532(0.652) | |||||||||
|
| ||||||||||
| CL (STL10) | CPTAC | −1.773(1.397) | ||||||||
| TCGA | −2.962(1.482) | |||||||||