Fig. 6

Heatmaps generated by unsupervised hierarchical clustering of the 22 selected features of VAT-bowel model in (a) training cohort and (b) the total validation cohort, and (c) examples of VAT and bowel feature maps overlaid on the CT images of four CD patients. The values of the heatmaps and feature maps are all standardized to the range of [0, 1], in order to achieve a clear view. In heatmaps (a and b), each row is one selected radiomics feature, and each column is a sample (PNR, red; PR, blue); the dendrogram at the top represents samples with similar information determined by clustering; the white arrows point to a VAT feature named “wavelet-LHH_glszm_LargeAreaLowGrayLevelEmphasis” or a bowel feature named “wavelet-LHL_glszm_LargeAreaEmphasis.” These two representative radiomics features are overlaid on CT images of four patients (cases a and b, with response to IFX therapy, predicted probabilities = 0.199 and 0.174; cases c and d, without response to IFX therapy, predicted probabilities = 0.622 and 0.523; VAT-bowel model’s cut-off value = 0.268) as shown in image (c). Both features demonstrate differences between patients in PR and PNR groups, with higher values from the PNR patients (cases c and d), suggesting more complex and coarse texture features of VAT and bowel. PNR, primary nonresponse; PR, primary response; IFX, infliximab; VAT, visceral adipose tissue; VAT-bowel model, radiomics model based on features extracted from VAT and the whole inflamed bowel