Fig. 4

Performance and representative results across different models on all lesions dataset. The highest performing model is CDGAM with a mean AUROC of 0.815 ± 0.032. The remaining performances are: 0.588 ± 0.047 (CLAM), 0.800 ± 0.038 (DAM), 0.810 ± 0.039 (DGAM), and 0.813 ± 0.041 (CDAM). CDGAM yields the heatmaps with highest segmentation and contrast quality. Note that CLAM highlights many tiles from glass whereas our CDGAM highlights liver tissues which contributes to the diagnoses.