Multi-scale texture analysis: from ¹⁸F-FDG PET images to pathological slides

Abstract

Purpose: Characterizing tumor heterogeneity using texture indices (TI) calculated from Positron Emission Tomography (PET) images has shown promise in predicting treatment response or patient survival in some types of cancer. Yet, the relationship between PET-derived TI, precise tracer distribution and biological heterogeneity needs to be clarified. We investigated the relationship between PET-derived TI and observations made on autoradiography and histological slides. Methods: Three mice bearing orthopically implanted mammary tumors derived from transgenic MMTV-PyMT mouse were scanned with the Inveon PET/Computerized Tomograph after injection of ¹⁸F-FDG. Tumors were then sliced, autoradiography was performed and tumor slices were stained using haematoxylin and eosin. Six TI computed from PET, autoradiography and histological images were compared to assess the ability of texture analysis to capture heterogeneity at different scales. Results: TI were significantly correlated (Spearman coefficient R between 0.57 and 0.85) between PET and autoradiography images but the TI values differed in magnitude. The TI correlation was low between histological and PET or autoradiography images (R between 0.06 and 0.54). All TI were little or moderately influenced by the difference of voxel size and spatial resolution in autoradiography images. TI measured in autoradiography images were significantly different (p<0.05) between histological regions with a high density of cells and regions with a low cell density, or between regions presenting different spatial arrangements of cells. Conclusion: Heterogeneity computed in vivo from PET images accurately reflects the heterogeneity of tracer uptake directly computed ex vivo from autoradiography images. Various tumor cell density and cell spatial distribution measured on pathology slides can be distinguished using TI calculated from autoradiography images despite the difference in voxel size and spatial resolution. Yet, tumor texture as assessed from PET images only coarsely reflects the spatial distribution and density of tumor cells.