Multiple Instance Learning with Center Embeddings for Histopathology Classification

Dual attention multiple instance learning with unsupervised complementary loss for COVID-19 screening

“…(2019) 0.803 0.796 0.767 0.85 0.742 DeepAttentionMIL Ilse et al. (2018) 0.859 0.875 0.861 0.75 1 JointMIL Chikontwe et al. (2020) 0.901 0.909 0.896 0.85 0.968 Zhang3DCNN Zhang et al.…”

“…(2019) 0.845 0.852 0.836 0.8 0.903 DeepAttentionMIL Ilse et al. (2018) 0.845 0.859 0.845 0.75 0.968 JointMIL Chikontwe et al. (2020) 0.845 0.837 0.814 0.9 0.774 D A -CMIL (w/o ) 0.718 0.728 0.714 0.65 0.806 D A -CMIL (w/o ) 0.873 0.88 0.866 0.825 0.935 D A -CMIL (w/ ) 0.958 0.955 0.951 0.975 0.935 …”

“…(2018) , ClassicMIL Campanella et al. (2019) and JointMIL Chikontwe et al. (2020) .…”

Scale-Aware Transformers for Diagnosing Melanocytic Lesions

“…Experimental results show that SA-Transformer-Network outperforms state-of-the-art methods by a significant margin. For example, SA-Transformer-Network is 8% more accurate than the method proposed by Chikontwe et al [7] and 6% more accurate than the method proposed by Hashimoto et al [8]. Importantly, SA-Transformer-Network delivers comparable performance to 187 practicing pathologists who interpreted the same test set cases in an independent study.…”

“…ChikonMIL. The method of Chikontwe et al (Chikon-MIL) (R3 in Table 2) [7] first selects the top-k patches, and then uses these patches for instance-and bag-representation learning. This method also uses a center loss that reduces intra-class variability and a soft assignment to learned diagnostic centroid for final diagnosis.…”