Introduction of a cascaded segmentation pipeline for parametric T1 mapping in cardiovascular magnetic resonance to improve segmentation performance

Abstract

The manual and often time-consuming segmentation of the myocardium in cardiovascular magnetic resonance is increasingly automated using convolutional neural networks (CNNs). This study proposes a cascaded segmentation (CASEG) approach to improve automatic image segmentation quality. First, an object detection algorithm predicts a bounding box (BB) for the left ventricular myocardium whose 1.5 times enlargement defines the region of interest (ROI). Then, the ROI image section is fed into a U-Net based segmentation. Two CASEG variants were evaluated: one using the ROI cropped image solely (cropU) and the other using a 2-channel-image additionally containing the original BB image section (crinU). Both were compared to a classical U-Net segmentation (refU). All networks share the same hyperparameters and were tested on basal and midventricular slices of native and contrast enhanced (CE) MOLLI T1 maps. Dice Similarity Coefficient improved significantly (p < 0.05) in cropU and crinU compared to refU (81.06%, 81.22%, 72.79% for native and 80.70%, 79.18%, 71.41% for CE data), while no significant improvement (p < 0.05) was achieved in the mean absolute error of the T1 time (11.94 ms, 12.45 ms, 14.22 ms for native and 5.32 ms, 6.07 ms, 5.89 ms for CE data). In conclusion, CASEG provides an improved geometric concordance but needs further improvement in the quantitative outcome.

Figure 1

Processing pipelines for refU, cropU and crinU. Convolutional neural networks (CNNs) are used for the segmentation of the myocardium as tissue of interest. While refU directly uses the input image, cropU and crinU use the region of interest image section that belongs to the 1.5 times enlarged bounding box from an object detection algorithm (ODA). In contrast to cropU, crinU uses a two channel image with the second channel having the original predicted bounding box mask.

Figure 2

Example results of the object detection algorithm showing bounding boxes for the left ventricular myocardium. The upper block corresponds to native and the lower block to contrast enhanced data; respectively in each block the first row corresponds with respect to the Dice Similarity Coefficient (DSC) and the second row corresponds with respect to the Hausdorff Distance (HD) while the first column shows the best and the second column the worst case. Green denotes true positive, blue false negative and red false positive segmented bounding box pixels.

Figure 3

Example results of the automated segmentation in refU, cropU, crinU and cropU_A. The first column shows the original image, the second column the refU segmentation, the third column the cropU segmentation, the fourth column the crinU and the fifth column the cropU_A segmentation. The upper block corresponds to native and the lower block to contrast enhanced data; respectively in each block the first row shows a fairly good case across all four pipelines, the second row shows a case that is improved in cropU and crinU compared to refU and the third row shows a poor case across all four pipelines. Green denotes true positive, blue false negative and red false positive segmented pixels.

Figure 4

Geometric results of the automated segmentation. The first column shows the geometric results for refU, the second column for cropU, the third column for crinU and the fourth column for cropU_A. The upper block corresponds to native and the lower block to contrast enhanced data; respectively in each block the first row shows the boxplots of the Dice Similarity Coefficient (DSC) and the second row shows the boxplots of the Hausdorff Distance (HD).

Figure 5

Quantitative results of the automated segmentation. The first column shows the quantitative results for refU, the second column for cropU, the third column for crinU and the fourth column for cropU_A. The upper block corresponds to native and the lower block to contrast enhanced data; respectively in each block the first row shows the correlation plot including the linear regression and the equivalence margin whereas the second row shows Bland–Altman-plots including the limits of agreement. Blue dots represent cases within the equivalence margin while red dots represent cases exceeding the equivalence margin.

Figure 6

Coherence analysis of the automated segmentation. The first column shows the coherence analysis for refU, the second column for cropU, the third column for crinU and the fourth column for cropU_A. The upper block corresponds to native and the lower block to contrast enhanced data; respectively in each block the first row shows histograms of disjoint segmented pixel values of the expert ground truth and the pipeline model and the second row shows the correlation plot between Dice Similarity Coefficient (DSC) and the absolute T1 error including the linear regression. Blue dots represent cases within the equivalence margin while red dots represent cases exceeding the equivalence margin.