Deep learning based automated left atrial segmentation and flow quantification of real time phase contrast MRI in patients with atrial fibrillation

Abstract

Real time 2D phase contrast (RTPC) MRI is useful for flow quantification in atrial fibrillation (AF) patients, but data analysis requires time-consuming anatomical contouring for many cardiac time frames. Our goal was to develop a convolutional neural network (CNN) for fully automated left atrial (LA) flow quantification. Forty-four AF patients underwent cardiac MRI including LA RTPC, collecting a median of 358 timeframes per scan. 15,307 semi-manual derived RTPC LA contours comprised ground truth for CNN training, validation, and testing. CNN vs. human performance was assessed using Dice scores (DSC), Hausdorff distance (HD), and flow measures (stasis, velocities, flow). LA contour DSC across all patients were similar to human inter-observer DSC (0.90 vs. 0.93) and a median 4.6 mm [3.5-5.9 mm] HD. There was no impact of heart rate variability on contouring quality (low vs. high variability DSC: 0.92 ± 0.05 vs. 0.91 ± 0.03, p = 0.95). CNN based LA flow quantification showed good to excellent agreement with semi-manual analysis (r > 0.90) and small bias in Bland-Altman analysis for mean velocity (-0.10 cm/s), stasis (1%), and net flow (-2.4 ml/s). This study demonstrated the feasibility of CNN based LA flow analysis with good agreements in LA contours and flow measures and resilience to heartbeat variability in AF.

Keywords: Atrial fibrillation; Cardiac magnetic resonance imaging; Deep learning; Image processing; Real time phase contrast.

Fig. 1

LA RTPC plane is placed parallel to the mitral valve. 2D RTPC raw data is reconstructed into magnitude and phase cardiac data based on acquisition time (366 frames total for this patient). The left atrium is contoured to quantify blood flow velocities. LA- left atrium

Fig. 2

RTPC magnitude volume data (2D image data + time) is preprocessed and inputted into a dense U-net architecture convolutional neural network. The network features 2,4,6,8, 12 dense (outputs of each layer feed forward as inputs to subsequent) layers per layer with each layer consisting of batch, relu, convolution and drop out with constant feature map size of 12. Max-pooling was utilized for down sampling within the encoder structure and transposed convolution within the decoder structure. The final layer consisted of a 1 × 1 × 1 convolution and softmax function to generate a probability of each voxel (2D spatial + time) of the magnitude input image volume belonging to the left atrium or the background for each time frame

Fig. 3

Best (A), median (B), and worst (C) dice score performing frames over the entire patient cohorot. Semi-manual (blue) is diplayed over AI (red) contours. Each frame comes from a different patient

Fig. 4

Flow curves for the patients having the best (top), median (middle), and worst (bottom) dice score frame. AI (red) displayed excellent agreement with Semi-Manual (blue) flow values over time, despite low performing frames

Fig. 5

Bland Altman analysis of agreement between semi-manual and AI per heartbeat measurement of peak velocity, mean velocity, net flow, and stasis