Abstract
163
Introduction: Gray level intensity within FDG-PET provides considerable information that could be linked to tumor biology. The aim of this study was to define image features of heterogeneity via radiomics analysis of radio-isotopic FDG images and histopathology offered by H&E staining of triple negative breast cancer (TNBC) patient-derived tumor xenografts (PDX), with a view to establish correlation between FDG-PET and H&E pathological prognostic factors in invasive breast cancer.
Methods: TNBC PDX were injected with ~100µCi FDG (N = 13). Sixty minutes post FDG injection, tumors were excised, and flash frozen for slicing. Approximately thirty-six 10 μm slices were obtained from each. For select tumors, alternate slices were used for histology or autoradiography. Tumor FDG spatial uptake intra-tumor heterogeneity was assessed by strict binary tree based iterative partitioning method and radiomics analysis. The earth mover distance (EMD) was used to measure the dissimilarity between the segmented regions within tumor. An efficient graph and normalized cut based approach were used to segment tumor into two regions. The optimum dividing line was chosen where voxel intensity difference between two regions was maximum. The heterogeneity was calculated from average EMD of the first three levels of strict binary tree. The robust radiomics histogram (10), GLCM (21), GLRLM (13), GLSZM (13), and NGTDM (N=5) features were extracted from 3D as well from 2D tumor images. Gray level quantization 64 was used for all higher order features. All volume dependent radiomics parameters were normalized to make tumor volume independent. A fast K-means method was implemented to visualize the different regions of the tumor based on heterogeneity.
Results: The proposed segmentation was tested and compared with pathology images for heterogeneity similarity, as shown in Figure 1(D-K). There was significant similarity (P value < 0.05) in radiomics features on corresponding regions between PET and H&E (Figure 1 (M)). The value of heterogeneity correlates well (P value < 0.05) with the optimal number of cluster determination (Calinski Harabasz) method (Figure 1 (L)). EMD using divide and conquer; and combination of Cluster Prominence, Maximum Probability, Dissimilarity, and Strength parameters from radiomics can be used as powerful features for heterogeneity assessment. The region scale plot of intra-regional EMD (shown in Figure 1 (C)) was used to compare the intra-tumoral heterogeneity between two multiple tumors.
Conclusions: Validation of intratumoral heterogeneity has potential to add analytical information to breast cancer treatment response, aggressiveness, and staging. The proposed scheme was used to quantify and validate intratumoral FDG-PET uptake against the underlying pathology, albeit at high resolution compared to clinical imaging. The strict binary tree based EMD heterogeneity analysis was used to establish a crucial relation between pathology and FDG images. Future studies will correlate intratumoral heterogeneity in FDG-PET imaging and correlation to pathology across scales.