A Machine Learning Model for Medical Image Recognition Using Texture-Based Features

摘要

This research proposes a supervised machine learning model to recognize human thyroid tissue from OCT images using texture-based features. The proposed model includes five procedures: 1) create training samples, 2) extract region of interests (ROIs), 3) measure texture features, 4) reduce feature space, and 5) train a supervised machine learning model. ROIs are obtained from original whole images. Then, texture properties of each image are extracted using 43 features, which can be summarized in four aspects: 1) statistic moments, 2) gray level co-occurrence matrix, 3) gray level run length matrix, and 4) gray level size zone matrix. To refine the feature space, four feature selection algorithms and six feature extraction models are applied to perform dimension reduction. After that, six classifiers are applied and tested to perform model selection. The proposed OCT image detection model is tested using collected human follicle samples. The performance is measured using five metrics: specificity, sensitivity, accuracy, area under ROC curve, and the number of features selected through a 5-fold cross validation approach. The experimental results show that the proposed model can achieve around 98% accuracy and significantly reduce feature dimension. This study provides a potential of using artificial intelligence techniques for patient tissue automatic recognition.