Table 3.
Performance of Classification Model on the UCB-CRC and DREAM Datasets
| DE Accuracy Within UCB-CRC (n = 24) | DE Accuracy Within DREAM (n = 399) | Instance-Wise Accuracy | Non-DE Class Accuracy | DE Class Accuracy | |
|---|---|---|---|---|---|
| Raw meibography image | |||||
| Baseline results | 4.2% | 97.7% | 94.4% | 92.2% | 96.7% |
| Raw image + eyelid detection | |||||
| Baseline results | 4.2% | 92.5% | 91.6% | 87.5% | 95.7% |
| Raw image + eyelid detection + tarsal plate segmentation | |||||
| Baseline results | 16.6% | 97.2% | 92.8% | 92.5% | 92.9% |
| + Rescale DREAM | 41.7% | 91.4% | 86.6% | 81.3% | 91.9% |
| + Adjust random crop | 54.2% | 88% | 84.2% | 76.7% | 91.7% |
| + Rotation and fine-tune | 70.8% | 91.7% | 80.8% | 72% | 90.5% |
The results indicate that the ability of the classifier to distinguish between DE and non-DE samples from tarsal plate masks can be attributed more to the inherent characteristics of the MG features rather than being solely influenced by the dataset source. This also reinforces the value of the segmentation-based approach, with the results serving as a baseline comparison for models utilizing raw meibography images without standardization and those enhanced with eyelid detection.