Comparison of Extracting Tissue- and Image-based Characteristic Features for Machine Learning Prediction of Colorectal Polyp Malignancy

Detection and classification of precursor polyp via computed tomography colonography (CTC) has shown its significant clinical impact on management of colorectal cancer (CRC). In this study, we compared the classification performance of five machine learning (ML) models and one deep learning (DL) model. The five ML models utilized the Haralick feature (HF), Haralick measure (HM), extended Haralick feature (eHF), and extended Haralick measure (eHM), and one Affine tissue elastic feature with Karhunen-Loève (KL) Transformation on virtual monoenergetic images (Affine_KL VMI) that included 10 energies and were trained with a Random Forest classifier. The DL model was a 17-layer 3D convolutional neural network (3D_CNN) model including 3 instance normalization layers. The performance of the six models was evaluated using the area under the receiver operating characteristic curve (AUC).The AUC values for the six models ranged from 0.870 to 0.997 for 100 polyp masses with a diameter larger than 30mm, 0.711 to 0.985 for 182 polyp masses with a diameter between 10 to 30mm, 0.601 to 0.890 for 357 polyp masses with a diameter less than 10 mm, respectively. The Affine_KL VMI model consistently outperformed the other models, achieving the highest AUC values across all datasets, ranging from 0.890 to 0.997. This indicates that the Affine_KL VMI ML model is a promising approach for detecting and classifying precursor polyps in CTC images and may have significant clinical impact on the management of CRC.