International Journal of Applied Mathematics and Machine LearningAuthor's: Yang Hou
Pages: [23] - [40]
Received Date: February 24, 2018; Revised March 8, 2018
Submitted by: Omer Faruk Ertugrul
DOI: http://dx.doi.org/10.18642/ijamml_7100121932
For overcoming many challenges for lung cancer classification in CT image, such as (1) the size and shape of lung cancer in different parts of the body are uneven; (2) the locations of the edge of the lung cancer confirmed by different doctors (or the same doctor at different times) are different; (3) the same type of lung cancer in different patients has different growth directions; and (4) it is difficult to judge the classification effect etc. In this paper, we firstly extract a total of 5 categories of 385 radiomics features, and explain the significance of pathological organization of each type of features by the imaging principle. Then we select the logistic regression with L1 regularized model as our predictor. When the segmentation is not too fine for lung squamous cell carcinoma and adenocarcinoma, the classification also achieved a good result with an ACC of 75% and an AUC of 75.3%. To further illustrate the usefulness of these five broad categories of features in overcoming the difficulty of lung cancer imaging classification, we select a more detailed segmented lung cancer for classifying images of benign and malignant images, eventually we can get a better result with ACC of 85% and AUC of 90.7%.
radiomics features, lung cancer, classification, CT image, logistic regression, clinical explanation, L1 regularization.
