Spectral CT Inspired Data Engineering for Colon Polyp Classification

Spectral CT works with multiple energy X-ray sources or energy recognized detector, which can obtain the attenuation map under different energy. Therefore, spectral CT can enhance the contrast between soft tissues and enrich the textures of colon polyps. In this paper, inspired by spectral CT, we proposed a novel data engineering method, which could prominently enhance colon polyp classification by the enriched tissue textures. 63 polyp volumetric CT images (31 benign and 32 malignant) were obtained from a clinical CT scanner with effective 75keV X-ray energy. All polyps were resected with pathology reports serving as ground truth. A linear scaling was used to simulate the CT spectral data in Hounsfield units from the 75 keV original CT images to five energy channels from 35keV to 75keV based on the attenuation-energy response curve of three tissue types (i.e. fat, water and cellular tissue). We investigated the classification performance on the five channel spectral data by both deep learning or CNN and Random Forest (RF). Three types of inputs were considered, one is the spectral raw CT images, one is the co-occurrence matrices (CMs) extracted from the CT images, and the third one is the Haralick features (HF) extracted from the CMs. In accordance, three classification model was investigated based on two classifiers of RF and CNN. Experimental results showed the AUC (area under the receiver operating characteristic curve) score was improved by 12.6%, 0.44% and 3.3% for the spectral images, CMs and HFs respectively on the five-energy spectral data comparing to the original 75kev data only. Difference images scheme among neighboring energy channels will further improve the AUC score to 17.4%, 3.6% and 4.5% respectively. Moreover, the CM- and HF-inputs can achieve the best AUC of 0.934 and 0.927.