Session Item

Pre-operative radiotherapy and chemo-radiotherapy have been shown to reduce the risk of local recurrences after total mesorectal excision (TME) surgery. These neoadjuvant treatment strategies do not appear to improve survival and may negatively affect the quality of life after treatment. Hence, it is crucial to accurately select the patients who will or will not benefit from neoadjuvant therapy. Under-staging of the tumor and lymph nodes may lead to the omission of a beneficial pre-treatment, whereas over-staging may cause unnecessary morbidity. This study investigates the predictive value of radiomic features derived from pretreatment CT images for pathological lymph node (pN) staging for rectal cancer patients.

A retrospective study of 64 diagnosed colorectal cancer patients treated with short-course (5x5Gy) radiotherapy followed by total mesorectal excision (TME) surgery within ten days between 2007 to 2015.  The regions of interest (ROI) were delineated manually by an experience rectal radiologist. Radiomics features extraction was implemented using the Ontology-guided Radiomics Analysis Workflow (O-RAW) software. A total of 105 radiomic features extracted from each segmented ROI (Mesorectum and gross tumor volume of the primary disease (GTVp1) ) of pretreatment CT images were analyzed. Missing clinical information was imputed using the multivariate imputations by chained equations (MICE) package in R. Principal component analyses (PCA) was employed for dimensionality reduction after all features with a correlation coefficient above 0.7 have been excluded. Repeated (50) 5-fold cross-validation decision tree models are used to classify patients based on their pN status (Negative or Positive). The area under the receiver operating characteristic curve (AUC) is used to measures the performance of these models.

Figure 1 shows the developed trees from the meserectum and tumor data, respectively. The decision tree used three radiomics mesorectum information to make a decision. However, just two variables are used for the tumor information and one for the clinical.




Figure 2 shows the mean AUCs and confidence intervals of the trees discriminating abilities on the mesorectum, tumor, and clinical data are 0.62 (0.61-0.64), 0.59 (0.58-0.60), and 0.58 (0.57-0.60) respectively.

We observed that mesorectum radiomics features have a statistically significant higher discriminating ability compared to the tumor (p-value = 0.004) and clinical (p-value = 0.003) informations despite the lower than optimal prediction accuracy. However, there was no difference between the tumor and clinical data (p-value = 0.429). Our study highlights the potential benefit of a non-invasive and cost-effective radiomics for precision medicine, which could enhance the efficiency and efficacy of cancer care. The optimal performance of these models with further research and increased sample size is the next step for this project in addition to external validation.