FastVentricle: Cardiac Segmentation with ENet

Abstract

Abstract. Cardiac Magnetic Resonance (CMR) imaging is commonly used to assess cardiac structure and function. One disadvantage of CMR is that post-processing of exams is tedious. Without automation, precise assessment of cardiac function via CMR typically requires an annotator to spend tens of minutes per case manually contouring ventricular structures. Automatic contouring can lower the required time per patient by generating contour suggestions that can be lightly modified by the annotator. Fully convolutional networks (FCNs), a variant of convolutional neural networks, have been used to rapidly advance the state-of-theart in automated segmentation, which makes FCNs a natural choice for ventricular segmentation. However, FCNs are limited by their computational cost, which increases the monetary cost and degrades the user experience of production systems. To combat this shortcoming, we have developed the FastVentricle architecture, an FCN architecture for ventricular segmentation based on the recently developed ENet architecture. FastVentricle is 4× faster and runs with 6× less memory than the previous state-of-the-art ventricular segmentation architecture while still maintaining excellent clinical accuracy.

Multi-task Learning for Left Atrial Segmentation on GE-MRI

“…Motivated by Curriculum Learning [14], we trained our network firstly with cropped images where the left atrium taking a large portion of the image and then we gradually increased the image size. In this way, our network learns to segment from easy scenarios to hard scenarios and this helps the model to quickly converge in the beginning [15]. Despite the change in input images sizes, our network could still output a fixed length feature vector for classification since we employed spatial pyramid pooling [7].…”

Regional dynamics of fractal dimension of the left ventricular endocardium from cine computed tomography images

“…Additionally, fractal dimension analysis does not require manual contours to be drawn to define the LV endocardium and epicardium, which can be susceptible to inter and intra-observer variability 25 . With current advances in automatic cardiac segmentation [26][27][28][29] , it is reasonable to expect fractal dimension analysis to become fully automated. Since the method is also independent of motion tracking, it can be used to characterize topography even on a single static image of the heart.…”

“…While this is not ideal, we note that the aim of this work was to evaluate the feasibility of using topography variation as a surrogate measure of cardiac function, rather than to develop an optimized segmentation algorithm. Given the current advances in machine assisted segmentation techniques, such as convolutional neural networks [26][27][28][29] , we expect that the choice of threshold will very shortly not be subject to inter-user variability. While we have demonstrated the applicability of the method to images acquired with standard clinical protocols, the variability in CT image acquisition (different scanner models) and CT image reconstruction (different fields of view, slice thicknesses, and reconstruction kernels) of these images may introduce biases in the fractal dimension estimates.…”

ScarGAN: Chained Generative Adversarial Networks to Simulate Pathological Tissue on Cardiovascular MR Scans

“…The segmentation network is a U-Net-based DeepVentricle [13] with 96 initial filters, 4 downsampling layers, 3 convolutional layers before each downsampling or upsampling operations, and skip connection between the corresponding downsampling and upsampling layers. This segmentation network is first trained on 1143 steady-state free precession (SSFP) SAX CMR scans with per-pixel crossentropy as the objective function and optimized using Adam [12] with a learning rate of 1e −4 .…”