TABLE 5.
Impact of model components assessed through ablation on IU X-ray and VinDr-CXR datasets.
| Model | IU X-ray dataset | VinDr-CXR dataset | ||||||
|---|---|---|---|---|---|---|---|---|
| Accuracy | Recall | F1 score | AUC | Accuracy | Recall | F1 score | AUC | |
| w./o. Disentangled Temporal Representation | 88.10 0.02 | 85.37 0.03 | 86.14 0.02 | 91.34 0.03 | 86.20 0.02 | 83.77 0.02 | 84.51 0.03 | 89.94 0.02 |
| w./o. Ontology-Aware Alignment | 89.33 0.03 | 86.02 0.02 | 86.83 0.03 | 91.92 0.02 | 87.51 0.03 | 84.60 0.02 | 85.91 0.02 | 90.87 0.03 |
| w./o. Causal-Aware Refinement | 89.01 0.02 | 86.78 0.02 | 87.09 0.03 | 92.40 0.02 | 87.92 0.02 | 85.17 0.02 | 86.25 0.03 | 91.32 0.02 |
| Ours | 90.42 0.02 | 87.56 0.02 | 88.33 0.03 | 92.84 0.02 | 88.73 0.02 | 85.91 0.02 | 86.67 0.03 | 91.03 0.02 |
The values in bold are the best values.