Table 5.
Comparison of different image preprocessing methods
| Image preprocessing | Segment-level |
Document-level |
||||||
|---|---|---|---|---|---|---|---|---|
| Recall | Precision | F1 | AUROC (95% CI) | Accuracy (95% CI) | ||||
| AHI | Gray scale | 0.7187 |
0.9249 | 0.8089 |
0.9699 (0.9601–0.9796) | 95.45 (93.04–97.87) | ||
| Gray scale+dilate and erode | 0.6961 |
0.9687 | 0.81 |
0.9679 (0.9573–0.9784) | 94.41 (91.74–97.07) | |||
| Gray scale+contrast 20% | 0.7126 |
0.9324 | 0.8078 |
0.9705 (0.9609–0.9802) | 94.06 (91.32–96.80) | |||
| Gray scale+contrast 60% | 0.7268 |
0.9216 | 0.8127 |
0.9692 (0.9593–0.9790) | 95.45 (93.04–97.87) | |||
| Gray scale+dilate and erode+contrast 20% | 0.7315 |
0.914 | 0.8126 |
0.9743 (0.9652–0.9833) | 94.76 (92.17–97.34) | |||
| Gray scale+dilate and erode+contrast 60% | 0.7268 |
0.9216 | 0.8172 |
0.9715 (0.9620–0.9810) | 95.80 (93.48–98.13) | |||
| SaO2 | Gray scale | 0.7258 |
0.8617 | 0.7879 |
0.9334 (0.9190–0.9478) | 91.61 (88.40–94.82) | ||
| Gray scale+dilate and erode | 0.7427 |
0.8819 | 0.8063 |
0.9620 (0.9504–0.9736) | 90.21 (86.77–93.65) | |||
| Gray scale+contrast 20% | 0.6957 |
0.8889 | 0.7805 |
0.9431 (0.9296–0.9566) | 91.26 (87.99–94.53) | |||
| Gray scale+contrast 60% | 0.6863 |
0.8671 | 0.7662 |
0.9495 (0.9366–0.9623) | 91.61 (88.40–94.82) | |||
| Gray scale+dilate and erode+contrast 20% | 0.683 |
0.8871 | 0.7718 |
0.9523 (0.9398–0.9647) | 91.61 (88.40–94.82) | |||
| Gray scale+dilate and erode+contrast 60% | 0.6863 |
0.8671 | 0.7684 |
0.9486 (0.9356–0.9616) | 91.96 (88.81–95.11) | |||
Note: Each image preprocessing method was followed by fine-tuning a downstream ClinicalBERT. We highlighted the highest AUROC and Accuracy in bold.