Textural analysis of 18F-FDG PET/CT to predict tumor response of carbon-ion radiotherapy in patients with locally advanced pancreas cancer

3005 Objectives: Histopathological heterogeneity affects the responses of malignant tumors to therapies. Texture analysis can semi-quantify PET tracer distribution representing tumor heterogeneities. The present study aimed to determine whether the texture of advanced pancreatic cancer on 18F-FDG PET/CT images before carbon-ion radiotherapy (CIRT) can predict early treatment responses. Methods: We collected 18F-FDG PET/CT data from 37 patients with locally advanced pancreas cancer who were scheduled to undergo CIRT. The PET/CT images were analyzed using LIFEx software version 3.64 (lifexsoft.org). Volumes of interest (VOI) on tumors were delineated using a threshold of 40% of the maximum standard uptake value (SUVmax) in each lesion. The voxel intensities of each VOI were resampled using 64 discrete SUV between 0 and 10 using an absolute resampling method to calculate texture indexes. Global features derived from gray level histograms, and regional features of zone matrices, as well as run length and local features determined from gray level co-occurrence matrices were calculated using texture analysis. Conventional parameters such as SUV, metabolic tumor volume (MTV) and total lesion glycolysis (TLG) were calculated. We analyzed a total of 44 texture indexes using receiver operating characteristics (ROC) curves. The cutoff value to evaluate progression-free survival (PFS) after CIRT was determined from ROC curves. Early treatment responses on local lesion after CIRT were evaluated using the Kaplan-Meier method for all texture indexes. Results: The gray level-zone length matrix (GLZLM) of regional features had the highest area under the ROC curve (AUC). Twenty of the 44 texture indexes significantly differed (log-rank test) in terms of local prediction of treatment effects after CIRT. The texture index of the gray level run length matrix (GLRLM) of regional features and the gray level co-occurrence matrix (GLCM) of local features most obviously differed significantly. The statistically significant difference of the local prediction of treatment effects after CIRT was determined by setting a cutoff value of 4.7 (P < 0.005) in GLCM contrast, LGZE. Early treatment responses on local lesion after CIRT significantly differed when the cutoff was 0.8 (P < 0.005), 1.7 (P < 0.005), in GLRLM_RP, dissimilarity. Conclusions: Texture analysis of PET/CT images before CIRT was useful. Textures such as GLCM and GLRLM might be promising parameters with which to predict early responses to CIRT among patients with locally advanced pancreas cancer.