Table 2.
The diagnosis accuracy (AUC) of COVID-19 image-only, text-only and image-text disease classification experiments.
| Methods | Year | Ratio of training data | Retrospective studies | Prospective studies | ||||
|---|---|---|---|---|---|---|---|---|
| Image-only | Text-only | Image-text | Image-only | Text-only | Image-text | |||
| CLIP54 | 2021 | 1% | 87.5 | 75.6 | 88.6 | 58.7 | 65.3 | 69.9 |
| ConVIRT34 | 2022 | 1% | 88.1 | 86.4 | 88.8 | 59.6 | 66.4 | 71.5 |
| BioViL30 | 2022 | 1% | 90.4 | 89.7 | 91.0 | 60.9 | 68.8 | 73.0 |
| Med-MLLM | Ours | 1% | 95.3(0.3) | 93.8(0.5) | 95.9(0.4) | 64.8(1.1) | 72.9(0.8) | 78.2(0.7) |
| CLIP54 | 2021 | 100% | 95.7 | 83.3 | 89.0 | 63.5 | 68.8 | 75.2 |
| ConVIRT34 | 2022 | 100% | 97.6 | 94.5 | 97.7 | 70.4 | 77.6 | 82.1 |
| BioViL30 | 2022 | 100% | 97.4 | 94.5 | 98.2 | 66.7 | 80.5 | 84.4 |
| Med-MLLM | Ours | 100% | 98.4(0.2) | 96.3(0.4) | 98.7(0.2) | 81.0(0.4) | 84.1(0.5) | 90.3(0.3) |
All values are reported in percentage (%). The best results are in bold.