Multiple-instance learning with global and local features for thyroid ultrasound image classification

Multi-modality thyroid ultrasound image can provide more information about the lesion for the physician to diagnosis. In this paper, the thyroid B-mode ultrasound image and the elastogrom are viewed as a bag. And the local features of the B-mode image and the global features of the elastogram are considered as instances of the bag. Multiple-instance learning (MIL) method is employed to solve thyroid ultrasound image classification problem. Local features of B-mode are mapped to the concept space by self-organizing map (SOM). The hue component of elastogram is extracted to represent the elasticity information of the lesion. The bag vector is composed of the concept vector of the B-mode and global elasticity of elastogram. Finally, a traditional supervised learning method, support vector machine (SVM), is employed for classifying the lesion. The experimental results show that the proposed method can achieve better performance.