Fig. 7
From: Comprehensive Benchmark Dataset for Pathological Lymph Node Metastasis in Breast Cancer Sections
Impact of noisy data on model performance across different feature encoders. Bar plots show the AUC and F1-score of three MIL models (a) CLAM-MB25, (b) AMD-MIL33, and (c) FR-MIL37-when noisy data is added either to the training set or to the test set. Each group of bars compares the performance across multiple feature encoders, including ResNet-5026, ViT-S12, PLIP3, CONCH7, CONCH-V1.521, Ctranspath6, Gigapath5, UNI4, and Virchow14. Models exhibit more robust performance degradation when noise is added to the test set, while training set noise leads to more varied effects depending on the encoder.
