Attention-driven tree-structured convolutional LSTM for high dimensional data understanding

Abstract

Modeling sequential information for image sequences is a vital step of various vision tasks and convolutional long short-term memory (ConvLSTM) has demonstrated its superb performance in such spatiotemporal problems. Nevertheless, the hierarchical data structures (e.g., human body parts and vessel/airway tree in biomedical images) in various tasks cannot be properly modeled by sequential models. Thus, ConvLSTM is not suitable for analyzing tree-structured image data that has a rich relation among its elements. In order to address this limitation, we present a tree-structured ConvLSTM model for tree-structured image analysis which can be trained end-to-end. To demonstrate its effectiveness, we first evaluate the proposed tree-structured ConvLSTM model on a synthetic Tree-Moving-MNIST dataset for tree-structured modeling. Experimental results demonstrate the superiority of the tree-structured ConvLSTM model for tree-structured image analysis compared with other alternatives. Additionally, we present a tree-structured segmentation framework which consists of a tree-structured ConvLSTM layer and an attention fully convolutional network (FCN) model. The proposed framework is validated on four large-scale coronary artery datasets. The results demonstrate the effectiveness and efficiency of the proposed method, showing its potential use cases in the analysis of tree-structured image data.