Table 22 Cross-domain generalization evaluation. Significant values are in bold.
Dataset/modality | Cell Types | Imaging technique | Original performance | With dual-gland GAN | Improvement | p-value |
|---|---|---|---|---|---|---|
Fluorescence microscopy | ||||||
Human protein atlas | 17 types | Confocal fluorescence | 85.2% accuracy | 92.6% accuracy | + 7.4% | < 0.001 |
Cell painting dataset | 12 types | High-content imaging | 82.7% accuracy | 89.8% accuracy | + 7.1% | < 0.001 |
Allen cell collection | 8 types | Live cell imaging | 81.4% accuracy | 88.3% accuracy | + 6.9% | < 0.001 |
Brightfield microscopy | ||||||
DIC cellular Dataset | 15 types | Differential interference | 78.9% accuracy | 85.1% accuracy | + 6.2% | < 0.001 |
Phase contrast collection | 10 types | Phase contrast | 77.3% accuracy | 83.7% accuracy | + 6.4% | < 0.001 |
Specialized modalities | ||||||
Electron microscopy (EM) | 6 types | Transmission EM | 74.6% accuracy | 80.8% accuracy | + 6.2% | < 0.001 |
Multi-photon Dataset | 8 types | Two-photon microscopy | 79.2% accuracy | 85.9% accuracy | + 6.7% | < 0.001 |
Super-resolution (STORM) | 5 types | Stochastic optical | 76.8% accuracy | 82.4% accuracy | + 5.6% | 0.002 |
Cross-species validation | ||||||
Mouse cell atlas | 14 types | Confocal fluorescence | 80.3% accuracy | 87.1% accuracy | + 6.8% | < 0.001 |
Drosophila cellular DB | 9 types | Fluorescence imaging | 79.7% accuracy | 85.9% accuracy | + 6.2% | < 0.001 |
Yeast cell collection | 7 types | Brightfield/fluorescence | 75.4% accuracy | 81.2% accuracy | + 5.8% | 0.001 |
Clinical/pathological | ||||||
Cancer cell lines | 20 types | Various modalities | 83.1% accuracy | 89.7% accuracy | + 6.6% | < 0.001 |
Histopathology dataset | 12 types | H&E staining | 81.9% accuracy | 87.8% accuracy | + 5.9% | < 0.001 |
Average performance | 12.4 types | Mixed modalities | 79.8% accuracy | 86.4% accuracy | + 6.6% | < 0.001 |
