Table 4.
Detailed information on the segmentation studies.
| Author | Year | Type of method | Brain structure | Gestational age | Description of method | Learning strategy (if applicable) | US machine, US Probe, 2D/3D |
Number of subjects | Outcome |
|---|---|---|---|---|---|---|---|---|---|
| Al-Bander et al. | 2019 | DL | Fetal head | 10–40 | A FCNN was trained for segmentation and refined using ellipse fitting. |
Data-augmentation: cropping, rotation, zooming Annotations: during acquisition, trained medical researcher Strategy: no cross-validation, external test set |
Voluson E8 or 730†, ?, 2D |
335 | Dice = 0.98 |
| Gofer et al. | 2021 | ML | Fetal head, CP | 12–14 | Two segmentation algorithms were compared: 1) statistical region merging, which uses image intensities, and 2) trainable Weka segmentation, which is based on an ensemble of machine learning algorithms. Trainable weka segmentation performed best. |
Data-augmentation: none Annotations: two obstetricians with a subspecialty in fetal imaging Strategy: k-fold cross-validation, no external test set |
Voluson E10†, 6–12 MHz TVa, 3D |
56 | Mean percentage error: 1.71% |
| Gutierrez-Becker et al. | 2013 | ML | CER | 18–24 | A point distribution model was used. This model is a special case of statistical shape model. Thirty points were used to segment the cerebellum. |
Data-augmentation: none Annotations: expert fetal medicine specialist Strategy: 20-fold cross-validation, no external test set |
Voluson 730†, 4–8 MHz, 3D |
20 | Dice = 0.80 |
| Hesse et al. | 2022 | DL | CP, LV, CSP, CER | 18–26 | A 3D U-net for segmentation was trained using only 9 fully annotated volumes, combined with many weakly labeled volumes obtained from atlas-based segmentations. |
Data-augmentation: none Annotations: two experienced sonographers Strategy: no cross-validation, external test set |
Philips HD 9‡, TAb, 3D |
278 | CP = 0.85 LV = 0.85 CSP = 0.78 CER = 0.90 |
| Li et al. | 2020 | DL | Fetal head | 10–40 | A FCNN was trained, combined with ellipse fitting for the final segmentation. Simultaneously, fetal head measurements were performed with a special regression branch to regularize the segmentation result. |
Data-augmentation: Brightness, contrast, sharpness, Gaussian blur, flipping Annotations: during acquisition, trained medical researcher Strategy: no cross-validation, external test set |
Voluson E8 or 730†, ?, 2D |
335 | Dice = 0.97 |
| Moccia et al. | 2021 | DL | Fetal head | 10–40 | A CNN was trained to predict the HC distance field, bounding box, and segmentation of the fetal head. The CNN is based on a recurrent neural network, which is a specific type of architecture designed to propagate information across images. |
Data-augmentation: scaling, translation, rotation and shearing Annotations: during acquisition, trained medical researcher Strategy: no cross-validation, external test set |
Voluson E8 or 730†, ?, 2D |
335 | Dice = 0.98 |
| Shu et al. | 2022 | DL | CER | 18–26 | A U-net was combined with an adaptive soft attention module for segmentation. This attention module makes use of convolutional layers instead of fully connected layers. |
Data-augmentation: flipping, Gaussian blur Annotations: radiologist of the ultrasound department Strategy: no cross-validation, external test set |
Voluson E10†, 2.5–7 MHz Tab, 2D |
192 | Dice = 0.91 |
| Wu et al. | 2017 | DL | Fetal head | 19–40 | A cascaded FCNN was trained, which consists of multiple so-called levels of a FCNN. Every level uses information learned in the previous level. |
Data-augmentation: none Annotations: during acquisition, trained medical researcher Strategy: no cross-validation, external test set |
?, ?, 2D |
236 | Dice = 0.98 |
| Yaqub et al. | 2013 | ML | CP, CSP, CER, VC | 18–26 | A Random Forest was trained for segmentation. Distance features for the skull, the center of the head, and eye orbits were used besides classical image features. |
Data-augmentation: none Annotations: experienced clinician Strategy: no cross-validation, external test set |
iU22‡, ?, 3D |
20 | Dice: CP = 0.79 CSP = 0.74 CER = 0.63 VC = 0.82 |
Legend to brain structures: CER = cerebellum, CP = choroid plexus, CSP = cavum septi pellucidi, LV = lateral ventricles, VC = posterior ventricular cavity. 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. †GE Medical Systems, Zipf, Austria, ‡Philips, Bothell, WA 98021, USA.