Unsupervised Deep Learning for Hippocampus Segmentation in 7.0 Tesla MR Images

Abstract

Recent emergence of 7.0T MR scanner sheds new light on the study of hippocampus by providing much higher image contrast and resolution. However, the new characteristics shown in 7.0T images, such as richer structural information and more severe intensity inhomogeneity, raise serious issues for the extraction of distinctive and robust features for accurately segmenting hippocampus in 7.0T images. On the other hand, the hand-crafted image features (such as Haar and SIFT), which were designed for 1.5T and 3.0T images, generally fail to be effective, because of the considerable image artifacts in 7.0T images. In this paper, we introduce the concept of unsupervised deep learning to learn the hierarchical feature representation directly from the pre-observed image patches in 7.0T images. Specifically, a two-layer stacked convolutional Independent Subspace Analysis (ISA) network is built to learn not only the intrinsic low-level features from image patches in the lower layer, but also the high-level features in the higher layer to describe the global image appearance based on the outputs from the lower layer. We have successfully integrated this deep learning scheme into a state-of-the-art multi-atlases based segmentation framework by replacing the previous hand-crafted image features by the hierarchical feature representations inferred from the two-layer ISA network. Promising hippocampus segmentation results were obtained on 20 7.0T images, demonstrating the enhanced discriminative power achieved by our deep learning method.