Exploiting spatio-temporal information for view recognition in cardiac echo videos

Abstract

Abstract2D Echocardiography is an important diagnostic aid for morphological and functional assessment of the heart. The transducer position is varied during an echo exam to elicit important information about the heart function and its anatomy. The knowledge of the transducer viewpoint is important in automatic cardiac echo interpretation to understand the regions being depicted as well as in the quantification of their attributes. In this paper, we address the problem of inferring the transducer viewpoint from the spatiotemporal information in cardiac echo videos. Unlike previous approaches, we exploit motion of the heart within a cardiac cycle in addition to spatial information to discriminate between viewpoints. Specifically, we use an active shape model (ASM) to model shape and texture information in an echo frame. The motion information derived by tracking ASMs through a heart cycle is then projected into the eigen-motion feature space of the viewpoint class for matching. We report comparison with a re-implementation of state-of-the-art view recognition methods in echos on a large database of patients with various cardiac diseases.

Fine-tuned convolutional neural nets for cardiac MRI acquisition plane recognition

“…The previous work on cardiac view recognition has been concentrated mainly on real-time recognition of cardiac planes for echography (Otey et al 2006;Park et al 2007;Beymer and Syeda-Mahmood 2008). There exists some work on magnetic resonance (Zhou et al 2012;Margeta et al 2014;Shaker et al 2014).…”

A fused deep learning architecture for viewpoint classification of echocardiography

“…Progress on classification of viewpoints has been forged by a number of researchers [5][6][7][8][9] applying several approaches. For example, Ebadollahi et al [5] and Zhou et al [8] employ Markov Random Field models with a focus on spatial information to index echocardiogram (echo) videos through the detection of the number of objects presented on each image/frame (e.g.…”

“…4 chambers of the heart in A4C videos), giving rise to an averaged precision of 67.8%. In order to increase the classification accuracy, Beymer et al [6] take temporal information into account through the employment of scale-invariant feature points, achieving 80.1% of recognition rate. In their case, the extraction of motions is tracked by the application of Active Shape Models (ASMs) through a heartbeat cycle, which is then projected onto an Eigenmotion feature space of the viewpoint class for matching.…”

“…In addition, Otey et al [9] has proposed a different feature-based method that measures the magnitude of gradients in space-time domain of videos, which is built on a hierarchical classification scheme, in an attempt to reduce the number of misclassifications among super-classes. Since these features are hand-crafted, their application ranges vary depending on the characteristics of the feature , e.g., both work at [6] and [9] take only on four viewpoints into consideration, e.g. A2C, PLA, PSAP and PSAA at [6] and A4C, A2C, PLA, and PSA in [9].…”

“…Since these features are hand-crafted, their application ranges vary depending on the characteristics of the feature , e.g., both work at [6] and [9] take only on four viewpoints into consideration, e.g. A2C, PLA, PSAP and PSAA at [6] and A4C, A2C, PLA, and PSA in [9]. To take all eight viewpoints into consideration, the work detailed by Kumar et al [7] utilizes the technique of scale invariant features extracted from the magnitude image that has undergone edge filtered motion in advance, which is supported by Pyramid Matching Kernel (PMK) and Support Vector Machine (SVM) for view classification.…”