Introduction:
MRI plays an increasing and growing role in both clinical antenatal diagnosis and research, complementing Ultrasound screening for suspected fetal pathologies. While MRI offers higher spatial resolution, enhanced soft tissue contrast and a wide range of functional contrasts, it poses unique challenges, such as safety considerations as well as involuntary fetal motion and variability in fetal position and maternal surroundings. Specialist radiographers are thus essential. Recently, low-field 0.55T MRI, with the benefits of increased field homogeneity, larger bore size and longer T2* among others has been shown [1], carrying the potential to widen access to this modality. Radiological assessment requires high-quality images in defined planes for various measurements such as the bi-parietal diameter and trans-cerebellar diameter, to assess fetal growth [2]. While recent advances in Slice-to-Volume-Reconstruction (SVR) [3] allow 3D reconstructed high-resolution results and thus re-orientation to true brain anatomy, this technique is only available in specialist centres. Here, we present an automatic landmark detection and sequence planning, encompassing landmarking concepts from cardiac MRI [4]. Recent work highlighted the use of real-time AI methods in fetal MRI to perform quality control, automatic segmentation [5, 6] and automatic tracking [7]. Here, we present an automatic, Gadgetron-based fast landmark detection and subsequent automatic planning of true fetal brain scans, implemented, tested and evaluated on prospective 0.55T low field fetal MRI scans.

Methods:
The automatic sequence planning framework was implemented on a 0.55T scanner (MAGNETOM Free.Max, Siemens Healthcare, Erlangen, Germany). A whole uterus multi-echo gradient-echo single-shot EPI sequence in coronal maternal orientation (resolution=3.13–4.03, TE=[46/120/194/268/342]ms, slices=50–59, matrix size=100x100–128x128, resolution=3.13–4.03 mm), acquired to perform T2* mapping, was used subsequently to extract the fetal head position. The pipeline consists of two steps (illustrated in Figure 1). First, the EPI sequence was modified to export the data to a Gadgetron pipeline. A two-step deep learning-based detection using the nnUNet [8] framework in python first localises the fetal head and calculates the centre of mass (CoM) and then extracts specific head landmarks (fetal eyes and cerebellum) and stores the CoM and the patient coordinate system into a file. The first network was trained on 125 labelled fetal datasets acquired at 1.5T/3T and tested on 29 0.55T fetal datasets. The second was trained and tested on cropped images of 76/15 fetal subjects, respectively, for the eyes and cerebellum. The mean distance in millimetres between automatic and manual landmarks was calculated. Next, the following HASTE sequence was modified to use the stored information to calculate a plane crossing the midline of the fetal eyes and orthogonal to their connection line. Figure 1 illustrates a comprehensive schemata. The entire real-time pipeline was acquired prospectively in three pregnant volunteers (GA 34-38 weeks, MEERKAT study REC19/LO/0852) on the 0.55T MRI scanner. The resulting automatically planned sagittal planes were assessed by a radiographer and a radiologist with >15 years of fetal MRI experience answering the questions “Would you repeat this scan for planning purposes?” and “Rate the quality [0”unusable”-5”perfect”].” respectively.

Results:
The brain localization task, trained on mid/high-field data, achieved an overall DSC of 82.3±17.5% and IoU of 82.3±17.5% when tested on low-field scans across all TEs, fetal positions, and gestational ages. Regarding the landmarks, the mean distance between the centre of the cerebellum label for automatic/manual landmarks was 6.47 mm. Fig. 2 shows the predicted brain and landmarks segmentations for one fetal subject of the landmarks model test set. Fetal head landmarks were successfully localised and the HASTE sequences planned in three fetal scans. Fig. 3 displays the automatic planning results and manual radiographer-performed planning. Quantitative evaluation of the achieved sagittal planes resulted in one scan where a repeat would be required both for the automatic and the radiographer-planned attempts. Radiological scoring (1-5) was on average 3.0±0.0 for automatic and 2.33±1.27 for the radiographer-planned attempts.

Discussion and Conclusion:
Automatic planning of T2-weighted fetal MRI was demonstrated in three fetal subjects and comparison with manual planning revealed good agreement. Further planes and focus areas such as cardiac views will be addressed in the future. Planning sequences in fetal protocols based on information from the previously acquired images is novel and opens up the possibility of performing fetal MRI more widely in non-specialist centres.

Acknowledgements

The authors thank all pregnant women and their families for taking part in this study, the midwives Imogen Desforges, Chidinma Iheanetu Oguejiofor, and Maggie Lee for their invaluable efforts in recruiting and looking after the women in this study as well as Kathleen Colford, Kamilah St Clair, and Massimo Marenzana for their involvement in the acquisition of these datasets. This work was supported by a Wellcome Trust Collaboration in Science grant [WT201526/Z/16/Z], a UKRI FL fellowship, an NIHR Advanced Fellowship to LS [NIHR3016640], an EPSRC Research Council DTP grant [EP/R513064/1] and by core funding from the Wellcome/EPSRC Centre for Medical Engineering [WT203148/Z/16/Z]. The views presented in this study represent these of the authors and not of Guy’s and St Thomas’ NHS Foundation Trust.