Mr450

Fetal acquisitions were performed at 1.5T (MR450, GE Healthcare, Milwaukee, WI, USA) in the University Children's Hospital Zürich (KISPI) using a single-shot EPI sequence (TE = 63 ms, TR = 2200 ms) and 15 gradient directions at b = 700s/mm² (b700). The acquisition time was approximately 1.3 min per 4D volume. The in-plane resolution was 1x1 mm², the slice thickness was 4–5 mm, and the field of view 256x256x14−22 voxels. Three axial series and a coronal one were acquired for each subject. Brain masks were manually generated for the b0 (b = 0s/mm²) of each acquisition and automatically propagated to the diffusion-weighted volumes. Between 8 and 18, T2-weighted images were also acquired for each subject where corresponding brain masks were automatically generated using an in-house deep learning based method using transfer learning from Salehi et al. (45 ). Manual refinements were needed for a few cases at the brain boundaries.

Patients underwent prostate MRI on either a 1.5 T (15%) MR450 or 3.0 T (85%) Discovery MR750 HDx (GE Healthcare, Waukesha/Chicago, IL, USA) with a 16-32 channel surface phased array coil. The protocol has been described in detail previously¹² (link); in brief, this comprised axial Fast Spin Echo T1-weighted images of the pelvis, along with T2-weighted Fast Recovery Fast Spin Echo images of the prostate acquired in the axial (slice thickness 3 mm; gap 0-1 mm), sagittal and coronal planes (Table 2). Axial diffusion-weighted imaging (DWI) was performed using a dual spin-echoplanar imaging pulse sequence with slice thickness 3-4 mm; gap 0 mm, and automated apparent diffusion coefficient maps. An additional small-FOV (24 × 12 cm) DWI series was performed at 1400 s mm⁻² (1.5 T) or 2000 s mm⁻² (3 T). Dynamic contrast-enhanced (DCE)-MRI was acquired as an axial 3D fast spoiled gradient echo sequence following bolus injection of Gadobutrol (Gadovist, Bayer HealthCare) via a power injector at a rate of 3 mL s⁻¹ (dose 0.1 mmol kg⁻¹) followed by a 25 mL saline flush, injection at 28 s/40 s, temporal resolution 7 s/10 s. The axial T2, DWI, and DCE sequences were matched in orientation, slice thickness, and gap.

Imaging data used for training were acquired with the routine clinical protocol without any special preparation at 1.5T or 3T using body array coils on multiple MRI scanners (Architect, Artist, MR450, MR750, HDXT, GE Healthcare, Waukesha WI and Aera, Skyra, Vita, Siemens Healthineers, Erlangen, Germany). Pulse sequences acquired on the patients included axial T2, coronal T2, axial 3D spoiled gradient echo with either fat suppression or Dixon fat water separation, axial steady state free precession (SSFP), coronal SSFP and axial diffusion weighted imaging from mid-chest to below the bottom edge of kidneys. Because the axial T2 weighted images provided the best contrast between the organs and the background, they were routinely used at our institution for deriving organ volumes by manual contouring and have been used in prior deep learning studies of kidney segmentation [22 (link)]; this sequence was selected for the multiorgan segmentation. Axial T2 DICOM tag details are provided in Table S1.