Table 1 Prediction accuracy depending on hyperparameter values and feature encoders.

Learning rate	Weight decay	Metrics	ResNet	UNI	CONCH	CTrans-Path	Prov-GigaPath
0.0001	0.00001	Accuracy	0.7879	0.7500	0.7500	0.7273	0.7424
		B. Acc.	0.5531	0.6626	0.5776	0.6147	0.5883
		AUC	0.5524	0.6146	0.6546	0.5775	0.6149
		F1-score	0.8768	0.8370	0.8481	0.8252	0.8378
0.0001	0.0001	Accuracy	0.7879	0.7348	0.7348	0.7197	0.7197
		B. Acc.	0.5979	0.6130	0.5732	0.5998	0.6246
		AUC	0.5994	0.6304	0.6681	0.5627	0.682
		F1-score	0.8728	0.8328	0.8367	0.8189	0.8199
0.00001	0.00001	Accuracy	0.7576	0.7348	0.6818	0.6894	0.7121
		B. Acc.	0.5106	0.6343	0.5458	0.6243	0.5986
		AUC	0.5024	0.7172	0.559	0.5956	0.6549
		F1-score	0.8602	0.8291	0.7962	0.7844	0.8153
0.00001	0.0001	Accuracy	0.7576	0.7045	0.6970	0.6894	0.7273
		B. Acc.	0.5000	0.5819	0.5829	0.6243	0.6165
		AUC	0.5074	0.6650	0.5889	0.6193	0.6619
		F1-score	0.8614	0.8112	0.8039	0.7844	0.8251

The different cores per patient were first transformed into feature vectors using a ResNet-50 trained on ImageNet and different feature encoders trained on H&E-stained images: UNI by Chen et al. [29], CONCH by Lu et al. [30], CTransPath by Wang et al. [28] and Prov-GigaPath by Xu et al. [31]. The length of the feature vectors varies across the different feature encoders. The pretrained Resnet-50 and UNI generate 1024-dimensional features, CONCH produces features of size 512, CTransPath of size 768 and Prov-GigaPath outputs features of length 1536. B. Acc. stands for balanced accuracy. The results of our model are marked in bold.

Quick links

Search