Extended Data Fig. 10: Robustness of NEC under different segmentation accuracies.
From: Deep learning supported discovery of biomarkers for clinical prognosis of liver cancer
a, NEC scores calculated for each patient based on segmentation results generated by 11 CNNs. The NEC scores corresponding to ResNeXt50 (the CNN used in this study) are marked with an opaque blue asterisk. Patients are ranked based on NEC scores corresponding to ResNeXt50. b, Classification performance, segmentation results, NEC heatmaps, and prognostic performance of different CNNs. Histograms include recall, precision, and F1-score for each CNN’s ‘necrosis’ category tested on QHCG test set, as well as NEC prognostic performance (C-Index) based on segmentation maps generated by each CNN. c, Prognostic performance distributions of different CNNs (n = 11 networks). Boxplot whiskers extend to the smallest and largest value within 1.5 times the interquartile ranges of hinges, and box centre and hinges indicate median and first and third quartiles, respectively.