Cross-modality medical image detection and segmentation by transfer learning of shapel priors

Abstract

Machine learning based methods have been widely used for detecting and segmenting various anatomical structures in different medical imaging modalities. The robustness of such approaches is largely determined by the number of training samples. In practice it is often difficult to acquire sufficient training samples for a certain imaging modality. Since multiple imaging modalities are often used for disease diagnosis or surgical planning, images of the same anatomical structure may be available in a different modality. In this work we investigate the effectiveness of shape priors learned from a different modality (e.g., CT) to improve the segmentation accuracy on the target modality (e.g., MRI). The shape priors are exploited in the marginal space learning framework in several ways, e.g., increasing the pose hypothesis set, enriching the statistical shape model, and synthesizing new training images with real shapes. Experiments show that the additional shape priors transferred from a different source can dramatically improve the segmentation accuracy when the training set is small (e.g., with 10 or 20 training images).