Deep learning for colorectal cancer classification using dual energy CT

The objective is to predict the presence of budding, tumor stage, microsatellite instability (MSI) status, vascular/lymphatic permeation, peri-nervous sheathing, KRAS/BRAF mutations and grade of colon cancer on preoperative dual-energy CT imaging using radiomic analysis. This retrospective study consisted of radiomic analysis of preoperative dual-energy CT imaging of patients undergoing colon cancer resection. Radiologist Etienne Danse manually segmented the tumor region on dual-energy CT images. Pre-processing included resolution homogenization, correction of segmentations, and conversion of the RGB color images to a new color management system. 700 traditional radiomic features, 22428 traditional radiomic features after filtering (MM features) as well as 366864 deep features were extracted from the tumor region. Several prediction models were developed by varying the pre-processing method and the classifier used.The validation method was a partially nested cross validation. The performance of the models was evaluated using AUC, metrics (accuracy, F1, precision, recall), learning curve. From a total of 72 patients, 31 were segmented and 28 were finally selected in the final dataset. Most prediction models performed better with data pre-processing including segmentation correction and resolution homogenization. Most of the models did not require MM features and deep features. The models for prediction of budding and grade showed more than encouraging results (respectively, AUC of 0.92 with 80% accuracy and AUC of 0.97 with 90% accuracy). Preoperative prediction of the presence of budding, tumor stage, MSI status, vascular/lymphatic permeation, peri-nervous sheathing, KRAS/BRAF mutations and grade of colon cancer by radiomic analysis of the preoperative DECT scan adds specificity to the clinical assessment and may contribute to individualized treatment selection. In addition, the new color management system appears to concentrate information from the effective atomic number map into almost 3 times less data in some cases.