A Novel CT-Based Radiomic Analysis for the Prediction of Overall Survival in Rectal Cancer Patients Treated with Short Course RT Total Neoadjuvant Therapy

Abstract

Radiomics is a method based on the extraction of features from medical images, enriching the construction of predictive models for aiding patient management. The purpose of this study is to develop a model based on radiomic features for the overall survival (OS) in rectal cancer patients treated with total neoadjuvant therapy (TNT). Prediction of neoadjuvant rectal (NAR) score was also preliminarily evaluated. A retrospective cohort including patients treated with TNT at a single center between 2017-2021 was identified. TNT consisted of short course pelvic radiotherapy (25 Gy in 5 fractions) followed by FOLFOX or XELOX. A single CTV included mesorectum and nodal regions. Two endpoints were used: OS at 3 years (from the start of RT) and NAR score binary classification. A cut-off of 8.4 was used for NAR based on a previous local study. CTV radiomic features were extracted from simulation CTs using Pyradiomics v3. Features were extracted from the unfiltered images as well as from a set of eight filtered images. Radiomic features consist of 7 classes: shape, first order statistics, glcm, gldm, glrlm, glszm and ngtdm. 1967 initial features were extracted. Features with 0 standard deviation were removed as well as highly correlated ones (0.99 Pearson Coefficient). Dimension reduction using PCA was applied to each feature class, preserving 95% of the explained variance. Supervised algorithms were evaluated using a 10-fold cross-validation technique. Demographic and clinical features were also considered (age, sex, cT stage and N positivity) RESULTS: Eighty-two patients were included (37 females). Median age was 61.5 years (IQR 53-70). Clinical staging was comprised of 5 cT2 patients, 57 cT3 and 20 cT4, with 69 cN+. 3y-OS was 72%, with a median follow-up of 43.7 months (IQR 35.3-48.6). For the 3y-OS endpoint two prediction models were developed. First, a reference model (KNN classifying algorithm) based on demographic and clinical features only resulting in an AUC of 0.659 (SD 0.029). The second model (CART classifying algorithm) included demographic, clinical and radiomic features resulting in an AUC of 0.744 (SD 0.032). For the binary NAR classification model, a subset of 77 patients was used. Twenty-three presented a NAR<8.4 and 54 a NAR≥8.4. The CART algorithm presented the best performance with an AUC of 0.808 (SD 0.037) considering demographic, clinical and radiomic features. Radiomic models were developed for 3y-OS prediction and binary NAR classification. The addition of radiomic features to clinical and demographic information improved the predictive capability of the 3y-OS model. Radiomic prediction of NAR makes it a promising tool for the identification of potential short course RT TNT responders.