16. Predicting neoadjuvant therapy response in locally advanced rectal cancer using texture features

Abstract

Purpose To predict pathological response to neoadjuvant chemo-radiotherapy (CRT) in locally advanced rectal carcinoma (LARC), using a classifier based on texture features derived from MRI and PET acquisitions. Methods 47 patients with: (a) histologically diagnosis of LARC, (b) pre-treatment MRI and PET/CT, (c) neoadjuvant treatment consisting of 46–55 Gy in 23–30 RT fractions alone or in association with either infusional 5-FU or oral Capecitabine, and (d) total mesorectal excision were included in this study. Patients with complete (tumour regression grade, TRG = 1) or near complete (TRG = 2) regression were defined as responders (pR+), while patients with moderate (TRG = 3) to no regression (TRG = 5) were considered as non-responders (pR−) [1] , [2] . Before implementing texture analysis, tumours were semi-automatically segmented on T2-w MRI, ADC maps and PET/CT acquisitions. From the segmented tumours, the following quantitative features were extracted from T2-w, ADC and PET images: (a) first-order parameters: median, mean, percentiles (10th, 25th, and 75th), (b) SUV, metabolic volume and glycolytic volume only for PET images, and (c) 22 s-order texture parameters, derived from Haralick analysis [3] . Multivariate logistic regression was performed to identify features most correlated with TRG. Results Overall, 26 patients were classified as pR−, and 21 as pR+ after total mesorectal excision. Parameters included in the multivariate regression were “10th percentile PET”, “10th percentile T2-w”, “Homogeneity ADC”, “Homogeneity PET”, “Information measure of correlation T2-w” (p = 0.002). The area under the ROC curve was 0.83 (95% confidence interval = 0.69–0.93), sensitivity and specificity were 75% and 76%, respectively, in detecting responders. Conclusions Texture analysis could provide useful information in assessing response to neoadjuvant treatment in LARC patients. These preliminary results, if confirmed on larger dataset, could be useful to personalize the oncological pathway for patients, delaying or advancing surgery, according to the prediction of treatment response.