Table 2 TAF vs SARIFA: a comparison of adipose-associated histopathologic features/biomarkers
Stroma AReactive Invasion Front Area SARIFA | Tumor Adipose Feature TAF | |
|---|---|---|
Origin of concept | Guided by histopathologic experience, proposed and validated as prognostic H&E-based histopathologic biomarker with an underlying characteristic tumor biology | machine-learning derived image feature, that was consecutively identified by pathologists and further validated as potential biomarker |
Definition/description | SARIFA11,12,13,14:= area at the tumor invasion front, in which at least one tumor gland or a group of at least five tumor cells are directly adjacent to adipocytes without interjacent inflammatory infiltrate or desmoplastic stromal reaction | TAF19,26:= small moderately-to-poor differentiated tumor cell clusters adjacent to a substantial component of adipose tissue (very similar feature described recently27: predominantly adipose and inflammatory cells with occasional tumor cells) |
Location | Tumor invasion front | not applicable; identified on so-called tumor patches/tiles |
Investigated entities | Gastric, Colorectal | Colorectal |
kappa metric | in Colon Cancer: 0.87 and 0.77 in Gastric Cancer: 0.74 and 0.73 | in Colorectal Cancer: 0.69 (widespread vs others) |
Findings | SARIFAs are independently prognostic in gastric and colon cancer, and associated with other high risk features SARIFAs are associated with an upregulation of lipid metabolism and an altered immune response | TAF are independently prognostic in a binary and also in a semi-quantitative way TAF are the first validation of a biomarker initially extracted from a machine learning model and then validated by human pathologists |
Conclusion | SARIFA & TAF are both prognostic H&E image features/biomarkers that show similarities and overall are vivid and, as far as we know, first examples to really prove how histopathologic experience and machine learning can reinforce and benefit from each other. | |
