Denoised recurrence label-based deep learning for prediction of postoperative recurrence risk and sorafenib response in HCC

Table 2 The performance of various models

Model	1/4	1/3	1/2	2/3	3/4	1/1	Average	Parameters
ResNet18	0.999	0.946	0.926	0.885	0.873	0.724	0.892	44.60 M
DenseNet121	0.998	0.959	0.915	0.887	0.881	0.700	0.890	26.53 M
Swin-Transformer	0.997	0.953	0.938	0.900	0.897	0.704	0.898	334.82 M
CNN-SASM^a	0.997	0.975	0.951	0.916	0.901	0.737	0.913	46.94 M
CNN-SelfAttn^a	0.998	0.974	0.938	0.922	0.897	0.733	0.909	46.94 M
CNN-SelfAttn-Attn^a	0.997	0.970	0.941	0.913	0.902	0.743	0.911	47.06 M
CNN-CSA^a	0.998	0.972	0.949	0.909	0.898	0.743	0.912	42.92 M
CNN-SSA^a	0.997	0.965	0.932	0.887	0.888	0.718	0.898	46.90 M

^arepresents our proposed models in this study. All three combined CNN with self-attention mechanisms. CNN-SASM is the pathological feature extractor. CNN-SelfAttn was constructed by removing the softmax layer of CNN-SASM. CNN-SelfAttn-Attn was developed by substituting the softmax layer of CNN-SASM to the conventional attention layer. Both of them were applied to explore the role of the softmax layer. CNN-CSA is the CNN-SASM without the SSA module, and CNN-SSA is the CNN-SASM without the CSA module

ISSN: 1741-7015