Prediction of Complete Pathological Response to Neo-Adjuvant Chemoradiotherapy Using Magnetic Resonance Imaging-Based Radiomics Analysis in Locally Advanced Rectal Cancer.

Abstract

In locally advanced rectal cancer (LARC) patients treated with neoadjuvant chemoradiotherapy (CRT), surgery could be omitted in patients with complete pathological response (pCR) without compromising progression-free and overall survivals. Our objective is to develop a radiomic model based on Magnetic Resonance Imaging (MRI) and contrast-enhanced computed tomography (CT) scans to predict the pathologic complete response (pCR) to neoadjuvant CRT in LARC.All patients treated for a LARC with neoadjuvant CRT and subsequent surgery in 2 centers in the West of France (Insitut of Cancerologie Brittany Occidental (ICBO) and Nantes) were considered. Both pre-CRT pelvic MRIs and contrast-enhanced CT-scans were mandatory for inclusion. The tumor was manually segmented on the T2-weighted and diffusion axial sequences and on the contrast-enhanced CT-scan. Eighty-eight radiomic parameters (15 shape and geometry features, 11 first-order and 62 second-order) were extracted from each sequence using the in-house MirasÓ software, with a total of 1056 features by patient. The overall cohort was randomly split into two independent cohorts (training: 70% and testing: 30%). A strict feature set selection workflow based on the Spearman's correlation coefficient and the Area Under the Curve (AUC) was developed to reduce the number of features. Based on these selected features, three pCR prediction models (clinical, radiomics and combined: clinical + radiomics) were developed on the training set only with a random forest approach and a Bootstrap internal validation with n = 1000 replications. An optimal cut-off maximizing the model's performance was defined on the training set. Each model was then evaluated on the testing set, based on the AUC and the C-statistic calculated with the pre-defined cut-off. Finally, a posteriori harmonization using the ComBat approach was applied to account for imaging modalities heterogeneity.Of the 124 included patients, 14 had a complete response (11,3%). In the training set, the clinical model based on 2 parameters (initial T-stage and degree of tumor differentiation) obtained an AUC of 0.67 and a C-statistic of 0.65. The radiomic model based on 1 parameter (entropy histogram derived from T2 sequence) obtained an AUC of 1 and a C-statistic of 1. The combined clinical and radiomic model was strictly identical to the radiomics model and thus had the same performance. On the testing set, models resulted in a C-statistic of 0.70/0.73/0.73 for the clinical, radiomics and combined models, respectively. Finally, after harmonization, the radiomic model achieved a C-statistic of 0.77 on the testing set while the combined resulted in a C-statistic of 0 .75.Radiomic model based on T2-weighted pre-therapeutic MRIs sequences could help to predict pCR after neoadjuvant CRT in LARC.