Table 9 Compare with SOTA models.
From: Research on liver cancer pathology image recognition based on deep learning image processing
Models | Accuracy | AUC​​ | Parameters | Key innovation points | Major limitations |
|---|---|---|---|---|---|
​​Ours | 94.7 | 0.95 | 8.3 | – | – |
​​TransPath43 | 92.1 | 0.94 | 86.5 | Multi-spatial attention fusion mechanism | Sclerosing liver cancer classification performance to be improved |
​​Hover-Net44 | 91.8 | 0.93 | 35.2 | Cross-modal Transformer Pre-training | High computational resource requirement (> 24GB GPU) |
​​DeepLIIF45 | 93.4 | 0.94 | 41.8 | Multi-scale cellular topology modeling | Insufficient sensitivity to poorly differentiated cancers (only 84.3%) |
​​UNet3+46 | 90.5 | 0.92 | 28.7 | Multi-marker joint decoding | Dependence on immunohistochemistry-assisted labeling |
