Radiomics Model Based on Non-Contrast CT Shows No Predictive Power for Complete Pathological Response in Locally Advanced Rectal Cancer

Abstract

(1) Background: About 15% of the patients undergoing neoadjuvant chemoradiation for locally advanced rectal cancer exhibit pathological complete response (pCR). The surgical approach is associated with major risks as well as a potential negative impact on quality of life and has been questioned in the past. Still, there is no evidence of a reliable clinical or radiological surrogate marker for pCR. This study aims to replicate previously reported response predictions on the basis of non-contrast CT scans on an independent patient cohort. (2) Methods: A total of 169 consecutive patients (126 males, 43 females) that underwent neoadjuvant chemoradiation and consecutive total mesorectal excision were included. The solid tumors were segmented on CT scans acquired on the same scanner for treatment planning. To quantify intratumoral 3D spatial heterogeneity, 1819 radiomics parameters were derived per case. Feature selection and algorithmic modeling were performed to classify pCR vs. non-pCR cases. A random forest model was trained on the dataset using 4-fold cross-validation. (3) Results: The model achieved an accuracy of 87%, higher than previously reported. Correction for the imbalanced distribution of pCR and non-PCR cases (13% and 87% respectively) was applied, yielding a balanced accuracy score of 0.5%. An additional experiment to classify a computer-generated random data sample using the same model led to comparable results. (4) Conclusions: There is no evidence of added value of a radiomics model based on on-contrast CT scans for prediction of pCR in rectal cancer. The imbalance of the target variable could be identified as a key issue, leading to a biased model and optimistic predictions.

Pre-Treatment T2-WI Based Radiomics Features for Prediction of Locally Advanced Rectal Cancer Non-Response to Neoadjuvant Chemoradiotherapy: A Preliminary Study

“…The potential of radiomics features extracted from MRI T2-weighted images for predicting a pathological complete response of rectal cancer was demonstrated in several recent studies, which reported promising results of their radiomics models with AUCs ranging from 0.69 to 0.93 [ 51 , 52 , 57 , 69 , 70 , 71 ]. In contrast to MRI, a recent study had demonstrated that radiomics features extracted from CT images showed no predictive power for complete pathological response in LARC [ 72 ], while another research showed that MRI T2-WI radiomics model performed better than CT radiomics model for predicting the LARC response to nCRT [ 73 ].…”

“…The majority of the features included in our final radiomic score were obtained from filtered images using the wavelet filters. Wavelet filters are useful for signal denoising and there are several radiomics studies that applied them for different purposes [ 74 , 75 , 76 ], including in the field of rectal cancer [ 72 , 77 , 78 ]. In a recent research of He et al, which aimed to develop an MRI-based radiomics signature for tumor grading of rectal carcinoma, the most relevant features included in their classifier were derived from wavelet-filtered images [ 77 ].…”

“…Additionally, in the predictive model constructed by Liang et al for the prediction of metachronous liver metastasis in patients with rectal cancer, their optimal selected feature model based on T2-WI consisted only of parameters extracted from filtered images using wavelet filters [ 78 ]. Additionally, wavelet filters have proved to be useful even for preprocessing CT images, a significant proportion of the selected features from the recent study of Hamerla et al for the differentiation of complete LARC responders being obtained from wavelet-filtered images [ 72 ].…”

Radiomic signature of the FOWARC trial predicts pathological response to neoadjuvant treatment in rectal cancer

“…Multiple studies have recently applied radiomic analysis to predict pCR after neoadjuvant treatment in LARC patients [ 23 – 26 ]. However, previous CT-based models for predicting pCR after neoadjuvant treatment turn out to be controversial, which can be attributed to their retrospective design, the small size of cohorts, and non-contrast CT images that they used [ 27 , 28 ]. In addition, to our best knowledge, none of previous studies has evaluated the feasibility of combing CT-based and MRI-based radiomic signatures to predict pCR.…”

Radiomics in Oncology, Part 1: Technical Principles and Gastrointestinal Application in CT and MRI

“…Radiomics and Radiogenomics were also widely investigated as imaging biomarker in colorectal cancer (CRC) in the assessment of mutational status, nodal metastases, stratification patient risk, and in evaluation of response to therapy [ 76 , 77 , 78 , 79 , 80 , 81 ]. Regarding CRC Radiogenomics, Yang et al [ 82 ] retrospectively investigated whether CT-based radiomic signature could predict KRAS, NRAS, and BRAF mutations in CRC by analyzing a primary cohort (61 patients) and a validation cohort (56 patients).…”