Transfer learning drives automatic HER2 scoring on HE-stained WSIs for breast cancer: a multi-cohort study

Breast Cancer Research

Table 2 Comparison of models between our work and previous studies

Method	Training stage		Feature dimension	Aggregation method	Internal validation	External validation
Method	Feature extraction	Instance aggregation	Feature dimension	Aggregation method	AUC	AUC (mean ± std)
TL-PA	√	–	N/A	Rule-based	0.75	0.76 ± 0.01
TL-SlideGraph⁺	√	√	768	Graph-based	0.68	0.66 ± 0.08
TL-CLAM	√	√	768	Attention-based	0.57	0.64 ± 0.04
ResNet50-CLAM [16]	–	√	1024	Attention-based	0.74	0.60 ± 0.04
Farahmand et al. [12]	√	–	N/A	Rule-based	0.81^a	N/A
Rawat et al. [11]	√	√	512	MLP-based	0.71^a	N/A
DAB-SlideGraph⁺ [14]	√	√	4	Graph-based	0.75 ± 0.02^a	N/A
Valieris et al. [17]	–	√	1024	Attention-based	0.61 ± 0.01^b	N/A

std standard deviation, N/A Not available, DAB 3,3'-Diaminobenzidine, MLP Multi-Layer Perceptron, – No additional training required
^aThe best performance on TCGA-BRCA, as reported in the original paper
^bThe performance of M6 (included 5-level HER2 status data) on TCGA-BRCA, as reported in the original paper
The best performance for each method was highlighted in bold

ISSN: 1465-542X