Room: Davidson Ballroom B
Purpose: To investigate if pre-treatment imaging features and/or clinical characteristics can predict post-treatment response to chemo-radiation treatment in locally advanced rectal cancer (LARC); we further identified prognostic biomarkers that are best associated with post-RT clinical response.
Methods: Forty-five consecutive patients with LARC were included. Each patient underwent neo-adjuvant CRT with prescription dose of 50 Gy(25 fractions), followed by total mesorectal excision surgery after completion of RT. Each patient underwent longitudinal pre-, during and post-RT MRI scans, including T2 and diffusion-weighted MRIs, etc. Quantitative image features were extracted from pre-treatment T2 images for the gross tumor volume(GTV), resulting in a total of 2838 quantitative features, along with clinical information. Patient response was assessed by post-operative pathology or MRI and colonoscopy. We stratified patients into good- and poor-responder subgroups. A novel nonlinear dimensionality reduction technique called T-distributed Stochastic Neighbor Embedding was introduced to reduce high-dimensional tumor features to 182 features, then Gaussian kernel function was adopted to map the input space into a high-dimensional feature space; lastly the support-vector-machine model was utilized to create the predictive model. Five-fold cross validation method was utilized to train and validate the model. The model performance was evaluated using receiver operating characteristic (ROC) curve. We further identified the most important features (top 5%) that may affect prognosis.
Results: Among the patients, 54.7% patients responded well, while 45.3% achieved poor response. Pre-treatment image features predicted post-CRT response with a calculated AUC = 0.82. Nine patient specific features, i.e., shape/roundness, morphology/GLCM texture, intensity, as well as clinical N stage, etc, were identified as predominant predictors of CRT response for LARC.
Conclusion: Pre-treatment image features, in combination of patient characteristics, could be used for post-treatment outcome prediction for LARC. Systematic analysis of image features via longitudinal multi-parametric MRIs should lead to improved predictive value to guide personalized treatment.