Texture-based treatment prediction by automatic liver tumor segmentation on computed tomography

This study presents a novel approach to extracting discriminative texture features of a liver tumor in computed tomography (CT) scans, which are used to combine with medical records for survival prediction. The liver region is first located using an image segmentation method. A pre-learned tumor classifier follows to segment the tumors in the liver region. Next, two sets of feature points are detected: (1) feature points in the liver region; (2) randomly sampling several points in the tumor region. Using each feature point as the center of a region of interest (ROI), this study computes Gray-level Co-occurrence Matrix (GLCM) which is further used to derive the texture features of the ROI. Multiple ROIs and thus multiple texture feature vectors are derived in an input CT image. These textures are collected and clustered into four clusters, where each of them is represented by a representative texture feature vector. To further enhance the discriminative powder of the texture features, for each CT image, we select two representative texture feature vectors which are from the two clusters with the highest probability belonging to the tumor region and the liver region, respectively. The resulting tumor (liver) texture feature vector is then labeled as a positive (negative) example in order to train a tumor Support Vector Machine (SVM) classifier. In the diagnosis stage, the tumor SVM classifies an input texture feature vector and the classification result locates the liver tumors in a CT image. Given CT images of 72 patients, to associate the detected tumor texture feature vectors with the medical records, a treatment prediction dataset is constructed for mining the survival prediction model using logistic regression. In the testing stage, to input a tumor texture feature vector and the possible treatments to the survival prediction model, the system computes the survival probabilities and generates a treatment prediction report, which suggests the most suitable treatment for the patient. Experimental results show that the proposed method gives good performance in terms of the accuracy of survival prediction.