Fig 1.
From top to bottom: without annotation and with annotation in red. From left to right: first trimester with an HC of 65.1 mm (pixel size of 0.06 mm), second trimester with an HC of 167.9 mm (pixel size of 0.12 mm) and third trimester with an HC of 278.4 mm (pixel size of 0.24 mm). Note that the skull is not yet visible as a bright structure in the first trimester.
Fig 2.
Distribution of HC and GA for the study data.
The x-axis represents the GA that was estimated using the CRL. The y-axis represents the HC measured by the experienced sonographer.
Table 1.
Number of images in the training and the test set.
Fig 3.
Overview of the three evaluated quantification systems A, B, and C.
System A was optimized on training data from all trimesters. System B has two pipelines: pipeline 1 was optimized on training data from trimester one and pipeline 2 was optimized on training data from trimester two and three. System C uses three pipelines: pipeline 1, 2 and 3 were optimized on training data from trimester one, two and three, respectively. All pipelines of a quantification system are computed when the HC is measured in a test ultrasound image.
Fig 4.
Overview of the twelve Haar-like features utilized in the quantification system.
From top to bottom: 1. Edge features in horizontal and vertical direction (kernel size of two by two pixels). 2. Line features in horizontal en vertical direction (kernel size of three by three pixels). 3. Center-surround features (kernel size of three by three pixels). 4. Rectangle features (kernel size of two by two pixels). The left side of each row represents the features in upright direction. The right side of each row represents the features in rotated direction. The height and width of the features in rotated direction are larger compared to the upright direction, but they capture the same relationship between the neighboring pixels.
Fig 5.
A: Perfect pixel classifier likelihood map where only the fetal skull has a high probability (depicted in white) and the background a low probability (depicted in gray). The pixels outside of the FOV are depicted in black. The center detected by the hough transform is depicted in purple and the radial offset is depicted in green. This schematic example uses eight angles (Nangles) for the polar transform (depicted in blue). The sampling distance (Sdis) is depicted in red. B: The output of the polar transform. The dynamic programming algorithm is used to extract the shortest path from left to right.
Table 2.
Parameter sets for optimizing systems A, B, and C.
Table 3.
Final parameter settings of quantification systems A, B, and C after parameter optimization.
Fig 6.
Steps of quantification system C.
From left to right: pipeline 1, 2, and 3, respectively. From top to bottom: (1) Input image. (2) Input ultrasound image with overlay of pixel classifier likelihood ranging from green to red and Hough transform result in pink. (3) Polar transformed pixel classifier likelihood with overlay of dynamic programming in red. (4) Polar transformed ultrasound image with overlay of dynamic programming in red and repositioned dynamic programming result in green. (5) Ultrasound image with overlay of the highest five percentile repositioned dynamic programming pixels. (6) Ultrasound image with fitted ellipse in green and annotation of the experienced sonographer in red. In this example image, the pipeline that was optimized for the second trimester is automatically selected as the best result, since the edge of this fitted ellipse has the highest median pixel classifier output.
Table 4.
Results of the experienced sonographer (observer 1) compared to the classifier A, B and C and the medical researcher (observer 2) on the test set.
Table 5.
Mean difference with the reference GA (days) that was estimated using the CRL in the first trimester.
Fig 7.
The difference with the reference GA (days) that was estimated using the CRL in the first trimester.
Fig 8.
Results of quantification system C closest to the median ADF of system C.
From left to right: first trimester with a ADF of 1.8 mm, result second trimester with a ADF of 1.6 mm, result third trimester with an ADF of 4.2 mm and worst result first trimester with an ADF of 36.8 mm. From top to bottom: (1) The ultrasound image. (2) The ultrasound image with overlay of the pixel classifier likelihood ranging from green to red and the Hough transform result in pink. (3) The ultrasound image with an overlay of the highest fifth percentile repositioned dynamic programming pixels. (4) The ultrasound image with the fitted ellipse in green and the annotation of observer 1 in red.
Table 6.
Results of quantification system C for training and test set compared to observer 1.
Table 7.
Comparison of system C against the reported results published in literature.