Table 4.
The diagnosis accuracy of different methods on various diseases across CheXpert, NIH ChestX-ray, RSNA Pneumonia, SIIM-ACR Pneumothorax, and Shenzhen Tuberculosis datasets.
| Methods | Year | Ratio of training data | CheXpert | NIH ChestX-ray | RSNA | SIIM-ACR | Tuberculosis |
|---|---|---|---|---|---|---|---|
| ConVIRT34 | 2022 | 1% / 10% | 87.0 | 66.2 | 88.8 | 71.3 | 93.7 |
| BioViL30 | 2022 | 1% / 10% | 86.8 | 69.5 | 88.1 | 69.5 | 95.0 |
| REFERS32 | 2022 | 1% / 10% | 87.2 | 76.7 | 89.4 | 76.6 | 95.8 |
| Med-MLLM | Ours | 1% / 10% | 88.9(0.5) | 83.3(0.9) | 93.4(0.5) | 87.5(0.7) | 96.7(0.4) |
| ConVIRT34 | 2022 | 100% | 88.1 | 81.3 | 92.7 | 90.0 | 96.4 |
| BioViL30 | 2022 | 100% | 87.9 | 82.5 | 89.1 | 86.9 | 97.1 |
| REFERS32 | 2022 | 100% | 88.2 | 84.7 | 92.7 | 89.3 | 98.0 |
| Med-MLLM | Ours | 100% | 89.5(0.2) | 88.1(0.3) | 95.3(0.2) | 94.0(0.4) | 98.6(0.1) |
All values are reported in percentage (%). The best results are in bold.