Table 2.
Detailed information on the biometry studies.
| Author | Year | Type of method | Brain structure | GA | Description of method | Learning strategy (if applicable) | US machine, US Probe, 2D/3D |
Number of subjects | Outcome |
|---|---|---|---|---|---|---|---|---|---|
| Araujo et al. | 2014 | N | CM | 17–29 | 4D view† software, VOCAL function. | Voluson 730† ? 3D |
224 | ICC: 0.92 | |
| Bertucci et al. | 2011 | N | Vermian perimeter, cross-sectional area, and super inferior diameter | 18–35 | 4Dview† software. | External test set | Voluson 730 or E8†, 4–8 MHz TAb or 5–9 MHz TVa, 3D |
12 | Significantly smaller cross-sectional area (18-19w), and perimeter (28-29w) in abnormal cases |
| Birnbaum et al. | 2021 | N | Cavum veli interpositi, an interhemispheric cyst-like structure | 14–17 | 4Dview† software. | Voluson E6, E8 or E10†, 5–9 MHz TVa, 3D |
87 | TCD: 13.1–18.4 mm Cavum veli interpositi: 0.3–0.8 mm Detection: 45% |
|
| Budd et al. | 2019 | DL | HC | 18–22 | A U-net was used for segmentation, followed by ellipse fitting to determine the HC. Using Monte Carlo drop-out an ensemble of segmentation was obtained; cases with high variance for the HC estimation were rejected. |
Data-augmentation: flipping, rotation, Annotations: expert sonographers Strategy: no cross-validation, external test set |
?, ?, 2D |
540 | Error = 1.81 mm |
| Carneiro et al. | 2008 | ML | BPD, HC | All | Feature extraction of image regions that were segmented by a constrained probabilistic tree classifier. From the segmentation the measurements were derived. |
Data-augmentation: none Annotations: 15 expert sonographers Strategy: no cross-validation, 3 external test sets |
?, ? 2D |
1760 | Error: BPD = 2.73 mm HC = 8.34 mm |
| Cinar et al. | 2020 | N | Cavum septi pellucidi | 19–24 | 4D view† software, VOCAL software function. | Voluson E6†, 2–7 MHz TAb, 3D |
99 | ICC: Intermediate – experienced: 0.78 Novice – experienced: 0.50 Novice – intermediate: 0.57 |
|
| Grandjean et al. | 2018 | ? | BPD, HC | 17–29 | Smartplanes‡ software. | Resona 7‡, 5–8 MHz, 3D |
30 | Error: BPD = 4 mm, HC = 11 mm | |
| Hata et al. | 2012 | N | BV | 10–13 | 4D view† software, VOCAL software function. | Voluson E8†, 3.7–17.5 MHz TVa, 3D |
36 | ICC: 0.991 | |
| Pashaj et al. | 2013 | ML | BPD, OFD, HC | 11–40 | Syngo auto OB§ software. | Annotations: not mentioned | ? §, 2.5–6 MHz, 2D |
83 | Success rate: BPD = 79.89%, OFD = 81.80% HC = 85.97% |
| Pistorius et al. | 2009 | N | Ventricles of telencephalon, diencephalon, mesencephalon and rhombencephalon | 8–9 | 4D view† software. | Voluson E8†, 6–12 MHz TVa, 3D |
6 | Success rate: 66% for all ventricles | |
| Pluym et al. | 2021 | ? | BPD, HC, TCD, CM, LV | 18–22 | SonoCNS† Fetal Brain software. | Voluson E10†, 2–8 MHz Tab, 2D |
143 | ICC: BPD = 0.81 HC = 0.88 TCD = 0.50 CM = 0.23 LV = 0.26 | |
| Rizzo et al. | 2016 | ? | BPD, HC, TCD, CM | 19–22 | 5D CNS¶ software. | WS80A Elite¶, 1–8 MHz TAb, 3D |
120 | ICC: BPD = 0.974 HC = 0.995 TCD = 0.994 CM = 0.990 |
|
| Rousian et al. | 2013 | N | Brain ventricle fluid volume | 6–12 | BARCO I-Space VR system, V-Scope volume rendering software. | Voluson E8†, 4.5–11.9 MHz TVa, 3D |
112∗ | Success rate: 38% | |
| Ryou et al. | 2019 | DL | HC | 11–14 | 2D slices of the image were used as input for a multi-task FCNN which outputs the segmentation of head, embryo and limbs and classification of the plane. These steps were repeated for slices taken from all three views (coronal, sagittal and axial). To obtain the HC, ellipse fitting was used. |
Data-augmentation: none Annotations: checked by clinicians Strategy: no cross-validation, external test set |
HD9∗∗, V7-3, 2D |
21 | Error = 6.03 mm |
| Shehzad et al. | 2007 | ? | HC | 14–38 | Automatic ellipsoid mode software††. | EcoCee and Power Vision††, 3.0–4.2 MHz, 2D |
72 | Correlation = 0.9999, Mean: significantly different | |
| Sofka et al. | 2014 | DL | HC, BPD, OFD, LV, CM, CER | 16–35 | A sequential estimation and integrated detection network was used, which employs the spatial relationship between different measurements. This was used to guide training of the network to detect the HC. |
Data-augmentation: flipping. Annotations: 1 experienced sonographer Strategy: no cross-validation, external test set |
Antares and S2000§, ?, 3D |
107 | error: CER = 1.37 mm CM = 0.87 mm LV = 1.01 mm OFD = 2.31 mm BPD = 0.94 mm HC = 4.06 mm |
| Van den Heuvel et al. |
2018 | ML | HC | 10–40 | Haar-like features were extracted from the image and were used as input for a Random Forest to detect the fetal skull. The HC was extracted using the Hough transform, dynamic programming and ellipse fitting. |
Data-augmentation: none Annotations: during acquisition, trained medical researcher Strategy: three-fold cross-validation, external test set |
Voluson E8 or 730†, ?, 2D |
335 | Error: first trim. = 3.1 mm second trim. = 2.5 mm third trim. = 4.8 mm |
| Van den Heuvel et al. | 2019 | DL | HC | 15–40 | A two-step approach for minimum computational resource circumstances used the well-known VGG-net for detection of the head and U-net for HC estimation via segmentation |
Data-augmentation: flipping. Annotations: 1 experienced sonographer Strategy: no cross-validation, external test set |
SonoAce R3¶, ?, 2D |
39 | Error = 10.3 mm |
| Verwoerd-Dikkeboom et al. | 2008 | N | HC, BPD | 6–14 | I-space, a virtual reality system that uses a virtual pointer to measure length. | Voluson 730†, ?, 3D |
28∗ | ICC: BPD = 0.998 HC = 0.997 |
|
| Verwoerd-Dikkeboom et al. | 2010 | N | BPD, HC, OFD | 6–14 | I-space, a virtual reality system that uses a virtual pointer to measure length. | Voluson 730†, ?, 3D |
125∗ | Success rate: BPD = 96.8% OFD = 96.8% HC = 96.8% |
|
| Yazdi et al. | 2014 | DL | BPD, OFD | 19–25 | SonoBiometry† software. | Annotations: two experts, one resident and two students | Voluson E8†, ?, 2D |
95 | Error: BPD = −0.17 mm OFD = −0.06 mm |
| Zhang et al. | 2020 | DL | HC | 0–40 | Regression of features extracted by a CNN to predict the HC. The fetal head is not segmented explicitly. |
Data-augmentation: flipping, translation and rotation Annotations: during acquisition, trained medical researcher Strategy: 5-fold cross-validation, external test set |
Voluson E8 or 730†, ?, 2D |
199 | Error = 4.52 mm |
Legend to brain structures: BPD = biparietal diameter, CER = cerebellum, CM = cisterna magna, HC = head circumference, LV = lateral ventricles, OFD = occipitofrontal dimeter, TCD = transverse cerebellar diameter. Legend to description of method: CNN = convolutional neural network, FCNN = fully convolutional neural network; a brief explanation can be found in Supplementary Material 4. 2D = two-dimensional, 3D = three dimensional, a ∗ indicates longitudinal data. ICC = Intraclass Correlation Coefficient. †GE Medical Systems, Zipf, Austria, ‡Mindray, Shenzen, China, §Siemens, USA, ¶Samsung Medison, Korea, ∗∗Philips, Bothell, WA 98021, USA, ††Toshiba, Japan. TAb = transabdominal, TVa = transvaginal.