Automatic segmentation of ventricular volume by 3D ultrasonography in post haemorrhagic ventricular dilatation among preterm infants

Abstract

To train, evaluate, and validate the application of a deep learning framework in three-dimensional ultrasound (3D US) for the automatic segmentation of ventricular volume in preterm infants with post haemorrhagic ventricular dilatation (PHVD). We trained a 2D convolutional neural network (CNN) for automatic segmentation ventricular volume from 3D US of preterm infants with PHVD. The method was validated with the Dice similarity coefficient (DSC) and the intra-class coefficient (ICC) compared to manual segmentation. The mean birth weight of the included patients was 1233.1 g (SD 309.4) and mean gestational age was 28.1 weeks (SD 1.6). A total of 152 serial 3D US from 10 preterm infants with PHVD were analysed. 230 ventricles were manually segmented. Of these, 108 were used for training a 2D CNN and 122 for validating the methodology for automatic segmentation. The global agreement for manual versus automated measures in the validation data (n = 122) was excellent with an ICC of 0.944 (0.874-0.971). The Dice similarity coefficient was 0.8 (± 0.01). 3D US based ventricular volume estimation through an automatic segmentation software developed through deep learning improves the accuracy and reduces the processing time needed for manual segmentation using VOCAL. 3D US should be considered a promising tool to help deepen our current understanding of the complex evolution of PHVD.

Figure 1

A series of representative sagittal planes extracted from different 3D US of different patients. The contours were measured manually (in green) and using the automatic deep-learning approach (red) one slice at a time. The contours measured automatically were obtained by training the CNN with the manual labels. The large differences seen in several images explain the low value of the IoU metrics. The automatic contours are more accurate.

Figure 2

Temporal series acquired over several days of the left (L) and right (R) brain ventricle volumes of four patients (id1–id4). In total, the four series sum 69 3D US. The curves with squared markers are the volumes measured using the gold standard, VOCAL. The curves without markers are the measurements obtained with automatic segmentation. Using VOCAL, there are situations in which the ventricles cannot be measured like the left ventricle of patient id3.

Figure 3

Temporal series acquired over several days of the left (L) and right (R) ventricular volumes of six patients (id5–id10). In total, the six series sum 83 3D US. All of them were used to validate the accuracy of the trained SegNet CNN. The curves with squared markers are the volumes measured using the gold standard, VOCAL. The red curves (without markers) are the measurements obtained with automatic segmentation. Using VOCAL, there are situations in which the ventricles cannot be measured such as the left ventricle of patients id6 and id8, or the right ventricle of id9.

Figure 4

(a) The Passing Bablok regression line allows visual inspection of agreement between manual and automatic segmentation. The intercept represents constant differences while the regression line slope represents proportional differences among methods, none of which were significant. (b) The observed bias (difference (Y–X)) is 0.61 (SD 2.28) and 94% of the differences are included within the 95% agreement limit (see text for detailed interpretation).

Figure 5

Temporal evolution of ventricular index (VI), anterior horn width (AHW) and ventricular volume in one of the included patients. This patient was randomized to the high threshold intervention group for the ELVIS trial. The volume of the ventricles grows over time. On day 12, an intraventricular catheter was placed to allow CSF removal, allowing for a reduction of ventricular size.

Figure 6

Temporal evolution of ventricular of the ventricular system of one patient (id1) over a period of 22 days measured automatically using deep learning-based segmentation.

Figure 7

Representative coronal and sagittal planes extracted from a 3D US from a patient with PHVD after grade III GMH-IVH and right intraparenchymal haemorrhage. The contours of the ventricular system measured automatically are drawn in red.

Figure 8

Left, orthoslice visualization of 3D US of the head of one infant with PHVD. The dilated ventricles are clearly visible. Top right, example of 3D patch surface visualization of the ventricles obtained using deep learning. Bottom left, representative sagittal and coronal slices of one 3D US of one patient with PHVD. Bottom right, representative RGB image generated by stacking three consecutive sagittal slices, the corresponding labels, and the segmentation calculated using a convolutional neural network (CNN).

Figure 9

Left, description of the dataset and how it was distributed for training and testing the deep learning framework and for its validation against the gold standard (VOCAL). Right, the table shows the number of ventricles measured using VOCAL and/or the deep learning-based segmentation method and the intraclass coefficient (ICC) between both methods. Using VOCAL, there are situations in which the ventricles cannot be measured. For instance, in patients id3, id6, id8 and id9 the number of unmeasured ventricles reached 50% and the ICC is lower. Using deep learning, all ventricles from the ten patients could be measured.