Nature子刊:实时AI引导胎儿股骨MRI规划及生长自动量化
Real-time AI-guided planning of fetal femur bone MRI and automatic quantification of growth
研究团队在Nature Scientific Reports上发表了一项研究,开发了一个完全自动化的实时胎儿股骨MRI采集与测量管线。在0.55T低场强MRI系统上,该系统通过AI分割胎儿股骨并自动计算最佳扫描平面,显著提高了测量精度和效率。
Introduction
Fetal magnetic resonance imaging (MRI) plays an increasingly important role in complementing ultrasound (US)1 for prenatal diagnostics, providing detailed anatomical visualization and superior soft-tissue contrast, particularly in cases complicated by maternal obesity, oligohydramnios, or suboptimal fetal positioning2. While the main clinical indications for fetal MRI remain pathologies of the central nervous system (CNS)3, a wide spectrum of genetic syndromes and skeletal dysplasias also manifest with abnormalities in fetal bone development, many of which have prognostic and therapeutic implications4. Conditions such as thanatophoric dysplasia, osteogenesis imperfecta, campomelic dysplasia, and Jeune syndrome are characterised by shortened, bowed, or under-mineralised long bones5,6, often first suspected during routine prenatal screening2. Moreover, chromosomal anomalies such as trisomy 21, 18, and 13 may present with shortened femur or humerus lengths as soft markers7. Accurate prenatal assessment of long bones - particularly the femur - is therefore critical for diagnosis, monitoring, and counselling. It also contributes to gestational age estimation8, fetal weight prediction, and the detection of growth restriction9,10.
Ultrasound remains the first-line modality for fetal skeletal assessment, facilitated by its interactivity and real-time capabilities. However, in addition to known limitations regarding late gestation and varying fetal position, increasingly frequent abdominal adipose tissue and the scar tissue from previous Caesarean sections affect the quality of the acoustic window and subsequently impairs image quality11. However, situations such as breech presentation are not only associated with difficulties in obtaining reliable biometry, especially femur length12,13 and subsequently unreliable weight estimations, but also increased obstetric risk, rendering such predictions ever more important. MRI offers a valuable complement by providing a large field-of-view and consistent tissue contrast, independent of acoustic windows14. Yet, MRI-based femoral length assessment remains under-utilised in clinical practice, primarily due to technical challenges in image planning and alignment15,16. MRI has intrinsic limitations for bone visualization, including the absence of cortical bone signal due to its low proton density and rapid T2* decay, and its dependence on precise slice planning to capture long bones in-plane for quantitative measurement17. Without targeted planning, the fetal femur may be visualised only partially or in oblique sections, preventing reliable length estimations18. This is especially challenging in fetal imaging, where involuntary motion precludes long 3D acquisitions, making rapid 2D snapshot sequences the standard approach19.
The most commonly used sequence for fetal bone imaging, echo-planar imaging (EPI) with short echo times (TEs), is typically acquired in a 2D multi-slice approach to cover the region of interest20. Misalignment between slices often leads to oblique or fragmented bone representations, causing foreshortening artifacts and reducing measurement reliability21. As a result, the femoral diaphysis is frequently captured in cross-section or at an oblique angle, thus reducing diagnostic value. Accurate quantitative skeletal assessment thus relies on precise plane prescription - a process that is currently manual, time-consuming, and highly operator-dependent. Furthermore, distortions inherent to EPI sequences affect the achievable measurement accuracy and further underscore the importance of positioning near the isocenter to minimize distortion artifacts. Recently, fetal MRI at 0.55T has been re-introduced22,23 as a strategy to mitigate such distortions. The wider bore enhances comfort, particularly accommodating the pregnant anatomy, while the lower scanner, maintenance, and infrastructure costs improve accessibility. A barrier to wider accessibility, however, remains and is linked to the need for time-consuming manual planning and thus the availability of dedicated specialists. In the dynamic fetal environment, manual planning remains inefficient, subject to inter-observer variability, and prone to suboptimal slice positioning24. These challenges highlight the need for intelligent, real-time planning strategies that adapt to fetal pose and optimize acquisition geometry during the MRI examination.
AI techniques have demonstrated strong potential for automatization of fetal MRI workflows. Deep learning–based segmentation and motion-correction pipelines have enabled robust retrospective reconstruction and volumetric analysis of fetal structures21,25. Recent efforts have extended these capabilities to real-time applications: Hoffmann et al.26 introduced a rapid head-pose detection framework for automatic slice alignment in fetal brain MRI, while Neves Silva et al.27,28 proposed fully automated, real-time planning frameworks for both anatomical and flow imaging, demonstrating the feasibility of real-time prescription in clinical scanner environments.
In parallel, the computer vision community has seen major progress in object detection and instance segmentation with DETR-like architectures29. Mask-based Transformers represent objects using sets of learnable query embeddings, each associated with a segmentation mask30. Crucially, the number of such queries can exceed the number of target objects31, enabling a pool of alternative object representations. A bipartite matching procedure, typically implemented via the Hungarian algorithm, subsequently assigns predictions to ground-truth instances. Within medical imaging, Chen et al.32 adapted this formulation to volumetric input by redefining segmentation as mask classification with iterative transformer-based refinement. Their approach showed particular advantages for small or low-contrast anatomical structures.
Building on these advancements, we propose a fully automatic, real-time acquisition planning and measurement pipeline for EPI-based imaging of fetal bones, with a particular focus on obtaining in-plane acquisitions of the fetal femur. Our framework integrates rapid anatomical localization and AI-driven segmentation to identify long bones, estimate their orientations, and compute optimal scan planes in real-time. This approach is designed to be robust to fetal motion, reduce dependency on operator expertise, and ensure consistent femoral visualization across subjects. In this work, we detail the development, implementation, and validation of this automated planning tool in a cohort of fetuses scanned on a clinical 0.55T MRI system. We assess its performance in terms of alignment accuracy, time efficiency, image quality, and reproducibility of femoral measurements. By enabling reliable, real-time planning of fetal skeletal acquisitions, this method has the potential to expand the clinical utility of MRI in assessing fetal growth and diagnosing skeletal pathologies.
Methods
The complete real-time pipeline, graphically depicted in Fig. 1, is described in the following sections. In step 1, the proximal and distal ends of the ossified diaphysis are automatically identified; in step 2, the imaging plane optimally aligned with the femur is calculated in real-time and finally in step 3, the in-plane scan is acquired and the femur length (FL) measurement is automatically extracted.
Automated pipeline for MR-based fetal femur length assessment. Step 1: initial scout acquisition and femur localization; step 2: real-time computation of the optimal scan plane aligning the femur in-plane; step 3: adapted short echo time echo-planar-imaging (EPI) scan with femur in-plane, enabling accurate length measurement.
Step 1: Automatic identification of the femur end points
Accurate identification of femur endpoints is essential for reliable measurement of the femur bone length. We therefore employ a two-stage localization and segmentation pipeline designed to robustly detect the femur with regard to limited MRI contrast, progressing ossification associated with fetal growth and development across gestation and variable protocol parameters typically employed in fetal MRI.
A coarse region of interest is first obtained by using a small network to roughly predict the center coordinates of the fetal legs and the required cropsize. The network consists of four convolutional downsampling layers similar to a U-Net decoder33, followed by average pooling and a linear layer. Based on the predicted coordinates, a bounding box is extracted and the cropped volume is resampled to a standardized size of \(96 \times 96 \times 64\) voxels to improve consistency. Within the cropped volume, the femur is segmented using a 3D U-Net architecture followed by with a transformer-based refinement module inspired by TransUNet32 integrated into the decoder. To increase robustness to anatomical and contrast variability, the network learns a set of m femur representations, each encoded as a d-dimensional feature vector, to capture complementary characteristics of the target structure. In a coarse-to-fine manner, these representations are iteratively refined through attention-based interaction with selected input features from the U-Net decoder.
Token sampling
Transformer layers benefit from restricting cross-attention to informative spatial locations, particularly for small anatomical structures30,32. However, the inherently limited spatial resolution of EPI acquisitions results in discrete attention masks that are overly restrictive at deeper layers. To address this, an edge-aware token sampling strategy is introduced. Local structural variation is quantified using channel-wise finite differences34 along spatial axes to compute an “edgyness” score for each voxel in the U-Net feature map. After flattening the feature map, the k highest-scoring positions are retained,
$$\begin{aligned} \mathcal {T} = \{\textbf{x}_i \in \mathbb {R}^{C} \mid i = 1, \ldots , k\}, \qquad k < N , \end{aligned}$$
(1)
where N is the total number of tokens. Regions with higher “edgyness” scores are associated with larger local changes and therefore correspond to class or anatomical boundaries. This focuses attention towards informative, heterogeneous regions and limits the contribution of less relevant positions.
Mask generation via multi-representation aggregation
The refinement module iteratively transforms abstract queries into concrete, input-informed femur representations. Each refinement step corresponds to a transformer decoder layer consisting of cross-attention for query-feature interaction and self-attention as well as a feed-forward network to integrate the gathered information into each representation30. After the l-th layer, a lightweight classification head maps each representation to one of the n target classes, producing per-representation class probabilities. Each representation also predicts a candidate segmentation mask. The final voxel-wise probabilities are obtained by marginalizing across all representation–mask pairs35, weighting each mask by its class probability, i.e., their representational quality.
Soft optimal-transport mask matching
Training requires aligning the multiple predicted masks with the ground-truth segmentation. Rather than enforcing a rigid one-to-one correspondence via Hungarian matching, we employ a soft matching strategy based on entropy-regularized optimal transport (Sinkhorn algorithm)36. Let \(\mu \in \Delta ^N\) denote the ground truth object distribution and \(\nu \in \Delta ^M\) the distribution of the network’s confidence over the learned femur representations. Given a cost matrix \(C \in \mathbb {R}^{m \times n}\), where \(C_{ij}\) quantifies how well the representation i matches the object j, the optimal transport plan
$$\begin{aligned} \min _{T \in \mathbb {R}^{m \times n}_+} \sum _{i=1}^{m} \sum _{j=1}^{n} C_{ij} T_{ij} + \epsilon H(T) \end{aligned}$$
(2)
is computed subject to marginals \(T \textbf{1}_N = \mu\), \(T^\top \textbf{1}_M = \nu\). The entropy term H(T) thereby promotes smooth, fractional assignments, ensuring that multiple representations remain active rather than collapsing to a single dominant one. We set \(\epsilon = 0.1\) to balance smoothness and discriminative matching37. To mitigate class imbalance between background and femur voxels within each volume, the target distribution \(\mu\) is chosen to be non-uniform, assigning foreground objects ten times more mass than background, so that the transport plan allocates proportionally more mass to foreground matches.
Each cost matrix entry combines a mask similarity term with a classification term. Concretely,
$$\begin{aligned} C_{i,j} = \lambda _{mask}\, \mathcal {L}_{mask}(i,j) + \lambda _{cls}\, \mathcal {L}_{cls}(i,j), \end{aligned}$$
(3)
where \(\mathcal {L}_{mask}\) is a mask-quality loss given by a weighted combination of Dice loss and binary cross-entropy between the predicted and ground-truth segmentation, and \(\mathcal {L}_{cls}\) is a classification loss derived from the predicted class probabilities. The weights are set to \(\lambda _{mask} = 0.7\) and \(\lambda _{cls} = 0.3\) to emphasize accurate mask generation32.
Mask losses in both the matching and the final loss computation are calculated on K sampled points instead of the whole mask to avoid overwhelming background noise. During the matching process uniformly sampled points are shared across all predictions and ground truth masks, while in the final loss calculation a specific set of points is sampled for each query. Adapting the idea of importance sampling38, the indices from token sampling are re-used to guide loss computation towards critical regions. To ensure sufficient coverage \(K = 0.1 \cdot (HWD) = 58.982\) was chosen and edge-based points were augmented with randomly sampled points in a one-to-one ratio.
Echomix contrast augmentation
Most datasets used for training were acquired with multi-echo sequences. Since bone–tissue contrast varies substantially both across gestation due to progressive ossification and between echo times - and field strengths - due to T2* shortening, we propose Echomix, an augmentation strategy that enriches contrast variability while maintaining anatomical plausibility. During training, random crops from the second-echo volume are combined with the first echo via weighted summation.
File generation
Once the femur is localized and segmented, its endpoint and center coordinates are extracted, transformed into scanner coordinates, and stored in a file ready to be read subsequently by the MR sequence to compute the automated femur plane prescription (step 2). A separate file was also generated to record FL measurements performed on the initial scout scan, enabling later comparison with measurements obtained from the automated scan.
Fetal data acquisition
In this study, 217 datasets were included - 59 for training and 158 for testing - acquired from pregnant women between 18 and 40 weeks of gestational age. All participants provided informed consent as part of ethically approved studies (MEERKAT [REC: 21/LO/0742], NANO [REC: 22/YH/0210], MiBirth [REC: 23/LO/0685] and CARP [REC: 16/LO/1573]). All experimental protocols were approved by the corresponding Research Ethics Committees and conducted in accordance with applicable guidelines and regulations. Fetal MRI scans were performed at St Thomas’ Hospital between 2022 and 2024 using a 0.55 T clinical MRI system (MAGNETOM Free.Max, Siemens Healthineers, Erlangen, Germany) with a 6-channel body and a 9-channel spine coil. Multi-echo gradient-echo (MEGE) and diffusion-weighted EPI sequences were employed. All acquisition parameters are listed in Table 1. The fetal femur was manually annotated by JB and SNS focusing on the hypointense region.
Step 2: Planning of in-plane fetal femur planes
A whole-uterus EPI sequence, previously modified to support the acquisition of multiple echoes (MEGE)39, was further adapted to enable real-time, automatic plane prescription using specific coordinates computed from the detected fetal femur (step 1).
The FIRE framework40 was employed for real-time communication between the scanner and the AI networks. Processing was carried out on an external GPU-accelerated workstation (NVIDIA GeForce RTX 2080 Ti, NVIDIA Corporation), connected directly to the scanner computer. A subsequent EPI sequence with a single echo was further adapted to incorporate these coordinates for automatic prescription of an in-plane fetal femur scan during the preparation phase, replacing the otherwise manual planning step. The file generated in step 1 was therefore transferred in real-time from the workstation to the scanner host computer to be accessible to the sequence, and the stored coordinates were applied as follows: the two endpoints of the segmented femur were used to compute the femur plane, while the center point was used to define the slice stack center and field-of-view. In parallel, the image data acquired with the whole-uterus sequence was processed to extract real-time femur length measurements (see step 3). The self-prescribed scan retained all acquisition parameters of the above-described whole-uterus EPI sequence, except that only 10 slices were acquired using a single TE (57 ms), resulting in a total acquisition time of 4 s.
Step 3: Automatic length measurement
Following previously established ultrasound guidelines, the femoral length is defined as the distance between the proximal and distal ends of the femoral diaphysis, including the metaphyseal end plate41. In EPI acquisitions, this corresponds to the hypo-intense part of the bone. Each femur label is considered a set of 3D coordinates, where each point corresponds to a voxel belonging to the ossified femoral shaft. Since the femur exhibits - even in pathological cases - a predominantly straight and elongated morphology, the anatomical axis is approximated by a linear model obtained by Principal Component Analysis (PCA). The first principal component exhibits the direction of maximum variance in the data and thus aligns with the primary elongation of the femur. All points are orthogonally projected onto this principal axis, and the points with the minimum and maximum projection values are chosen as endpoints of the femur. Mimicking calliper-based measurements, the Euclidean distance between the extracted endpoints is set as the femur length (FL).
Retrospective evaluation
The proposed method was retrospectively evaluated on a total of 72 datasets acquired using the scanner setup and acquisition parameters described in Step 1 and 66 datasets acquired as part of the CARP study on a 1.5 T Philips scanner. The cohort included scans obtained at gestational ages between 18 and 40 weeks, comprising 53 healthy cases and 19 pathological cases, including pregnancies diagnosed clinically with fetal growth restriction, short long bones, and other anomalies. Healthy participants were specifically recruited from low-risk pregnancies, with no major adverse risk factors at booking, no abnormal findings at first-trimester combined screening, and, for pregnancies beyond 20 weeks, no significant abnormalities on the routine anomaly ultrasound. All MRI examinations were additionally reviewed by a radiologist, and only cases without pathological imaging findings were retained in the healthy cohort.
Manual FL measurements were performed for both femora by two experienced observers (M1, M2), blinded to each other’s segmentations. This allowed for a robust estimation of the error of the automatic method and an assessment of inter-observer variability. In 57 cases out of the 138 datasets, the MRI-derived FL measurements were additionally compared to clinical ultrasound biometry performed by experienced obstetricians. Given the small target size (\(\sim 0.1\%\) of total image voxels), the segmentation performance was evaluated using the Dice similarity coefficient (DSC) to quantify volumetric overlap and the 95th percentile Hausdorff distance (HD95) to assess boundary agreement.
The precision of the automatically extracted proximal and distal endpoints was quantified by measuring the Euclidean distance to manually annotated endpoints, as well as by evaluating the accuracy of the resulting FL measurements. Differences in FL were assessed using the mean absolute error (MAE) and mean absolute percentage error (MAPE) with respect to the manual measurements of M2. In cases where both femora were visible, the measurement consistency was evaluated by comparing the FL measurements of the left and right femur of same fetus. As the majority of calcification occurs during the third trimester of pregnancy42 and substantially affects the MR signal characteristics and fetal bone appearance, absolute and relative measurement error was further evaluated separately and compared between second-trimester (\(\le 27\) weeks) and third-trimester (\(>27\) weeks) datasets. Automatic and manual length estimations were correlated with gestational age using Pearson correlation, and deviations were quantified to assess the impact of GA on achieved precision. Kendall’s tau correlation analysis was furthermore performed to assess any potential correlations between MAPE and maternal body mass index (BMI).
To estimate the impact of oblique and cross-sectional views on measurement reliability, cases were divided into two groups based on the difference between MRI-derived and ultrasound-derived FL (\(\le 5\%\) and \(> 5\%\)) testing the hypothesis that larger deviations correspond to cases where the femur is split across multiple slice groups. Therefore, the number of slices covering the femur was quantified for each case, and group differences were statistically assessed using a two-sided unpaired t-test.
Real-time prospective evaluation
24 prospective datasets (Mean GA=36.1 weeks, [23–39 weeks]) were acquired on a 0.55T clinical MRI scanner with the acquisition setup described above. For all cases, ultrasound measurements were available, performed within 4 days (Mean 1.25 ± 1.03 days) of the fetal MRI examination. The complete pipeline was tested prospectively in these cases in real-time, starting with the initial localisation of the femur via a whole-uterus EPI scan acquired in the maternal coronal orientation, followed by in-plane automatically-planned femur acquisition and subsequent FL computation. Accuracy was evaluated by comparing the real-time automatically generated segmentations to manual endpoint labels, extracted post-scan. Femur measurements were calculated manually and automatically, and compared with the ultrasound-based measurements extracted by two obstetricians (1–2 years of experience).
Domain shift and robustness evaluation
Robustness and generalization ability was tested on 20 retrospective datasets with a significant domain shift: Change of the field strength, vendor and employed sequence with data acquired as part of the CARP study on a 1.5T Phillips MRI with a spin-echo diffusion weighted sequence. Each dataset consists of two acquisitions from the same day within a time interval of 30 minutes. Table 2 provides a detailed overview of maternal characteristics and obstetric factors for the retrospective and prospective cohorts. Bland-Altman plots were used to graphically assess the agreement between both EPI scans, 2) manual and automatic MRI-based measurements, 3) automatic MRI-based measurements and manual ultrasound-based measurements and 4) measurements on repeated acquisitions of the same fetus.
Ablation study
A series of ablation experiments was performed to access the contribution of the proposed adaptions on femur detection. All ablations were performed on the retrospective dataset described in subsection Retrospective evaluation and also analyzed against the U-Net as a baseline network. Sampling-based attention was compared against full as well as masked cross-attention and soft mask matching was evaluated against Hungarian matching and fixed index-based matching, where each target class is represented by a single representation. Evaluation was in line with the methodology applied so far based on number of correctly detected femur and average length error (MAPE). Length error was analyzed for 0.55T data sets with \(\le 27\) and a \(> 27\) weeks GA as well as for 1.5T data sets separately.
Results
The described pipeline was successfully implemented and evaluated in a total of 182 cases - 138 retrospective (0.55T: 72, 1.5T: 66), 24 prospective cases and 20 out-of-distribution cases. The overall istribution of cases across the training, validation, retrospective evaluation, and prospective evaluation cohorts are summarized in Fig. 2.
Overview of the data distribution across model development, retrospective evaluation, prospective testing, and domain-shift analysis. Data acquired at 0.55 T was used for model training, validation, and the main retrospective and prospective evaluations, whereas independently acquired 1.5 T data were used to assess generalization under a domain shift.
Retrospective validation
Following successful center-point detection and extraction of the relevant sub-volume in 136/138 cases, the femur was detected in 129 of these cases. Neither femur endpoint of neither right nor left femur could be detected in seven cases, corresponding to fetuses of a gestational age of approximately 18–22 weeks and were excluded from further analysis. In 114/138 cases, both femur bones were detected. The complete pipeline - including body localization, femur segmentation, and endpoint extraction - required an average of 1.9 s per volume [1.1–2.1 s] in offline testing mode. The femur segmentation achieved a Dice similarity coefficient of 0.66 ± 0.1 and a HD95 of 5.0 ± 2.9 mm relative to the manual reference segmentation.
The average spatial distance between manually and automatically identified femoral endpoints was 5.1 ± 3.2 mm [0.0–13.3 mm], with the maximum distance of 13.3 mm observed for a case with imaging artifacts in the femoral region. The resulting MAE between manual and automatic annotations was 3.3 ± 2.3 mm, corresponding to a MAPE of 5.1 ± 3.9\(\%\). In 88/114 cases the laterality of both femora was correctly classified. The mean length difference between the left and right femur was 2.7 ± 2.8 mm (4.2 ± 4.3\(\%\)) for manual measurements and 2.8 ± 2.7 mm (4.4 ± 4.4\(\%\)) for automatic FL estimations. The length difference between the left and right femur was significantly correlated with the slice count difference (\(r=.27, p <.05\)) as shown in Fig. 3A.
Evaluation of 72 0.55T retrospective cases. (A) Impact of the slice count difference on the length difference between both left and right femur. (B) Difference in femur length measurements between manual annotations by two observers (M1, M2) and the automatic method (A). (C) Overlay of automatic and manual endpoints on the image acquired at the first echo time.
The MAE of the cases acquired before 27 weeks of gestation (n = 8) was significantly lower compared to the \(>27\) weeks (n = 64) datasets (\(p=.037\)). However, there was no significant difference between both groups in terms of relative error (\(p=.565\)). In fetuses of gestational age \(\le\) 27 weeks (n = 8), the average difference was 2.0 ± 1.35 mm (5.2 ± 3.5 %), whereas in those > 27 weeks (n = 64), it was 3.2 ± 2.9 mm (4.8 ± 4.8 %). No difference in measurement error was observed between healthy cases (4.8 ± 3.9 %, n=53) and those with detected pathologies (4.6 ± 3.4 %, n=19). Both manual and automatic measurements of FL showed a strong, significant relationship (\(p<.001\)) with gestational age, \(r=.89\) and \(r=.92\), respectively. The Kendall’s tau correlation analysis revealed no significant relationship between measurement error (MAPE) and maternal BMI, \(\tau = -.02,~ p =.809\).
Figure 3B illustrates the results of inter-observer variability between the two manual observers, M1 and M2. Femur segmentations were performed by observer M2, and in all cases, manual and automatic measurements were derived from annotated endpoints. Absolute differences between M1 and automatic measurements ranged from 1.0 to 11.7 mm (mean: 6.2 mm), while differences between M2 and automatic measurements ranged from 0.5 to 7.8 mm (mean: 5.4 mm). Differences between the two manual observers ranged from 0.0 to 9.0 mm (mean: 4.9 mm). For six cases, the variability of endpoint annotations is depicted in Fig. 3C. Manual measurements performed on the EPI acquisition were on average 2.6 ± 3.9 mm (3.6 ± 5.5 %) smaller than corresponding ultrasound measurements, while the automatic estimations were reduced by 1.0 ± 4.1 mm (1.4 ± 6.1 %). Cases in which the MRI-derived FL differed by more than 5% from the ultrasound measurement showed a significantly greater number of slices spanning the femur (3.8 vs 7.3, \(p=.017\)) compared to cases with an error below 5%. Figure 4A presents for the entire retrospective cohort MRI-based FL (manual and automatic) against gestational age, as well as the matched ultrasound results when available, stratified into healthy and pathological cases (marked with a cross) and overlaid on normative curves from large cohort ultrasound studies43, whereas Fig. 4B illustrates the percentage based length estimation error across gestational age.
Quantitative results for the retrospective 0.55T cohort from 19 to 40 weeks of gestational age. (A) Comparison of automatic MRI measurements (green), manual MRI measurements (yellow), and ultrasound measurements (red), where available. Cases with pathologies are indicated by a cross, and ultrasound-based percentile curves are shown in gray. (B) Absolute percentage error of the automatic MRI-based length estimation relative to the corresponding manual measurement across gestational age.
Prospective validation
The described pipeline was successfully run prospectively in all 24 cases of the prospective cohort. Figure 5 presents a comprehensive visualisation of the real-time results obtained for four example cases in the prospective validation cohort, illustrating the EPI scan acquired in the coronal orientation of the uterus, the detected landmarks, the automatically re-planned acquisition, and the resulting FL measurement overlaid on the EPI image.
Illustrative results of the complete pipeline in 4 prospective cases. (Step 1) Landmark detection was performed on the initial scout scan and enabled full coverage of the femur in plane during the (Step 2) re-planned acquisition.
The detection of femoral landmarks and automatic planning of acquisition parameters took less than 6 seconds after completion of the first EPI sequence. On the initial EPI scan, the femur was successfully detected in all 24 cases and corresponding landmarks were extracted in real-time during the scan to automatically plan the following in-plane acquisition. Manual endpoint annotations were performed by M2 after the scans were completed. The complete elongation of the femur was captured in 22 cases, while automatically extracted endpoints were located along the femoral diaphysis rather than at the anatomical proximal and distal ends in two cases. Accordingly, the mean distance between automatic and manual endpoints on the out-of-plane scan was 6.1 ± 3.2 mm, resulting in a MAE of 3.8 ± 3.5 mm and a MAPE of 6.5 ± 5.8\(\%\) after exclusion of two segmentations covering less than 30\(\%\) of the bone. Manual measurements of the left and right femur bones of the same fetus deviated on average by 5.5 mm ± 6.2 mm. Automatic right-left FL measurements were not performed, as in most cases, only one femur was fully captured within the field-of-view and free from artifacts.
Of the automatically planned in-plane scans performed, all but one captured the femur in plane with both ends clearly visible. The second acquisition resulted in an MAE of 1.3 ± 0.4 mm and an MAPE of 2.3 ± 1.0\(\%\). The differences between manual and automatic measurements ranged between 0.7 and 1.8 mm, while automatic endpoints deviated on average 3.9 ± 1.7 mm from the manual annotation. In-plane length measurements differed from the initial estimation by 2.3 ± 0.9 mm for the automatic method and by 2.6 ± 0.4 mm for manual measurements. The mean deviation from ultrasound biometry was 1.5 ± 0.6 mm.
Domain shift and robustness results
In 19/20 cases acquired with a diffusion weighted sequence at least one femur was detected, while both femora could be detected in 15/20 cases. The mean difference in FL between the first and second scan was \(2.0 \pm 1.2\) mm for automatic and \(2.3 \pm 1.6\) mm for manual measurements. Automatic measurements differed on average by \(4.6 \pm 1.9\) mm from the manual reference.
Bland-Altman plots for robustness evaluation. (A) Absolute differences between initial and re-planned automatic MRI-based FL measurements, (B) absolute differences between automatic and manual MRI-based FL measurements, (C) absolute differences between automatic MRI and manual ultrasound measurements compared to gestational age and (D) absolute differences between two repeated scans for automatic and manual FL measurements.
The results of the robustness study are presented in Fig. 6, with a mean difference in FL of 0.2 mm between the first and second real-time acquisitions illustrated in the Bland-Altman analysis (Fig. 6A). The mean difference between automatic and manual FL equals 1.2 mm (see Fig. 6B), and between ultrasound and MRI 1.0 mm (see Fig. 6C). The difference between ultrasound and MRI-based automatic FL correlates with size (\(r=-.56, p<.05\)) but not with gestational age (\(p>.05\)). Measurements on repeated scans of the same fetus differ on average by 0.3 mm for both automatic and manual measurements (see Fig. 6D).
Ablation study
Average dice scores, length error (MAPE) and the number of cases with at least one detected femur are shown in Table 3 for 0.55 T datasets, divided in \(\le 27\) weeks and \(> 27\) weeks GA, and domain shift experiment. The type of cross-attention predominantly affects the performance in cases with \(\le 27\) weeks GA. Cross-attention with discrete attention masks increases the mean error in these cases to 7.3% and the number of failure cases from nine to 17. For segmentation of small femur targets both soft and fixed matching result in lower length error across all cases compared to Hungarian matching.
Discussion
We presented a fully automatic real-time pipeline for quantitative fetal femur evaluation - consisting of fetal femur segmentation, landmark extraction, re-acquisition in the most suitable plane, and length estimation. The approach detects femoral endpoints in an initial scout scan and subsequently plans an in-plane acquisition in less than a minute, thus reducing operator workload and facilitating standardized MRI-based skeletal assessment. The MRI-based femur length complements ultrasound-based measurements and allows assessment even in cases where precise ultrasound quantification is complicated due to maternal habitus, late gestation and complicated fetal position among others.Its larger field of view makes it helpful when evaluating suspected skeletal dysplasia or congenital limb abnormalities, because the femur can be interpreted in the context of the pelvis, spine, other long bones, and associated organ abnormalities. Although recent work typically acquired fetal bone scans at 1.5T or 3T20, the here presented method was shown on a 0.55T MRI scanner, where approximately 40\(\%\) longer T2* relaxation times contribute to the capture of the rapidly decaying signal from the bones more effectively. However, the methodology developed in this study is not limited to the application at 0.55T as demonstrated with the domain shift experiment.
Retrospective evaluation demonstrated reliable landmark detection with a mean endpoint localization error of 5.1 mm [0.0–13.3 mm] and FL errors of 3.3 mm [0.0–14.6 mm] (4.8 \(\%\)), with the relative error in femur length remaining consistent across gestational age and comparable between healthy and pathological cases. The MRI-based FL corresponds with normal ranges from ultrasound in the literature43, in particular when accounting for the inherent variability of gestational age estimations. Importantly, the proposed framework extends beyond previous approaches that primarily perform retrospective biometry on already acquired images by integrating anatomical localization directly into the acquisition process. Prospective testing confirmed the robustness of this AI-guided planning strategy, as even partial detection of the femoral shaft was sufficient for accurate automated plane prescription: while scout scans were often affected by artifacts, the automatically planned acquisitions successfully captured the femur in-plane. This prospective re-acquisition resulted in higher measurement accuracy, with in-plane estimations achieving a MAPE of 2.3 \(\%\) compared with 5.1 \(\%\) for out-of-plane measurements. These findings demonstrate the added value of combining automatic anatomical localization, plane prescription, re-acquisition, and quantitative measurement within a single pipeline, rather than relying on retrospective analysis alone. Notably, automatic measurements also showed fewer implausible reference percentiles than manual annotations, underscoring their robustness despite these limitations. Manual annotation variability, however, remained high: small differences in caliper placement translated into several millimeters of discrepancy, particularly in oblique or cross-sectional views.
Although the achieved Dice score of 0.66 indicates only moderate alignment, the amount of overlapping regions appears to only partially influence measurement accuracy. Compared to previous work on AI-driven biometry, the accuracy achieved falls within the reported range of MRI-based brain measurements44,45. The demonstrated in-plane results were comparable to automatic ultrasound biometry, with mean FL errors of 0.9–2.7 mm reported46,47. In line with literature, performance remains constrained by the rather large voxel size of the employed EPI sequences and present susceptibility artifacts.
Mask quality is a well-recognized factor limiting the performance of mask-classification architectures48, and it may further exacerbate the challenges associated with small-object detection30. While previous studies have demonstrated the benefits of allowing multiple queries to converge toward the same target31,32, masks corresponding to smaller targets are more susceptible to segmentation errors, leading to a higher rate of misclassifications during the early stages of training. Under the strict one-to-one matching commonly used in DETR-like approaches, gradient updates originate only from a subset of matched queries based on the similarity with the assigned class. As a result, frequent early misclassifications can hinder the specialization of queries representing small targets, potentially driving unstable or random assignments. This pattern aligns with the observed training instability when applying Hungarian matching to femur segmentation as well as the superior performance of more stable matching strategies (Table 3) and is consistent with previously reported convergence and stability issues in related architectures49,50. In contrast, the soft one-to-many matching strategy employed here reduces sensitivity to early classification errors and facilitates concurrent optimization of multiple representations of each target. Empirically, this approach not only enhanced training stability but increased generalization to unseen 1.5T datasets compared to a baseline without redundant representations.
Growing evidence suggests that MRI can provide more accurate estimations of fetal weight51 than the traditional Ultrasound-based Hadlock method52, based on the occipitofrontal diameter, abdominal circumference and the femur length. However, MRI-based approaches typically rely on volumetric methods, which complicates direct comparison with classical methods. Previous studies have reported a high degree of agreement between manual measurements of standard biometry parameters in MRI and ultrasound24 and particularly demonstrated the feasibility of the automatic assessment of cranial measurements in MRI53. An extension of the proposed method to measurements of the fetal head and abdomen could offer an MRI-based alternative for fetal weight estimation and allow a comparison between volumetric and classical approaches.
Recent animal experiments suggest, that fetal growth restriction may be associated with impaired chondrogenesis and decreased endochondral ossification54, but previous results in humans are inconsistent at 1.5T18. Due to its consistent tissue contrast and capabilities to image cartilaginous tissue, MRI-based skeletal assessment has the potential to detect subtle changes to identify epimetaphyseal features specific for placental insufficiency, which may be particular promising for currently hard to diagnose cases of late-onset fetal growth restriction. Key strengths include the robustness of the proposed method, carefully designed with additional region-of-interest detection and transformer-based steps to ensure reliable and generalizable performance for clinical applications. Even in cases where the entire femur was not fully captured, sufficient information was captured through partial bone fragments to enable subsequent planning and re-acquisition. The low latency achieved demonstrates the real-time feasibility of real-time implementation, as carefully demonstrated in the nested prospective study. Multi-observer assessments and comprehensive evaluation of all pipeline steps - from landmark detection to quantitative measurements - pave the way for clinical translation.
However, this study has several limitations. First, although the presented cohort is large for a prospective fetal study, expanding it would allow more comprehensive testing of the robustness and accuracy of automatic femur length estimations, particularly in the re-planned acquisitions. The gestational age range, while spanning from 19 weeks to term, is skewed toward later pregnancy weeks due to the opportunistic recruitment through a study focusing on birth. Future studies could target earlier stages of pregnancy more specifically.
Reduced amniotic fluid volume was not an exclusion criterion. Such cases may exhibit lower overall contrast and reduced separation between fetal structures, potentially complicating femur localization in scout acquisitions. While reduced fetal mobility may limit motion-related image degradation, closer apposition of the limbs to adjacent anatomy could affect plane prescription. As performance was not stratified by amniotic fluid volume, its independent effect could not be quantified.
The ultrasound biometry datasets used as the reference standard in this study, a requirement due to the limited MRI-based literature, carries the drawback of the widely acknowledged variability. Furthermore, recent work also suggests a systematic bias, with the femur length in MRI EPI-based measurements typically shorter compared to ultrasound24. The discrepancies may stem from distinct physical properties underlying ultrasound and MRI such as acoustic shadowing or partial-volume effects in lower resolution EPI. These factors hinder further evaluation of the automatic method and emphasize the need for MRI specific growth charts for femur length.
Conclusion
The successful integration of a complete pipeline for real-time, automatic femur length measurement in fetal MRI - demonstrated with robust and precise results across 151 participants - paves the way for future clinical integration and more individualised and accurate assessment of fetal growth.
Data availability
The data acquired and/or analyzed during the current study are not publicly available due to ethical, legal, and privacy restrictions associated with sensitive human subject data. Access may be granted on reasonable request, subject to approval by the relevant institutional review boards and data protection regulations. The code used in this study is available from the corresponding author upon reasonable request.
Code availability
The code used in this study is available from the corresponding author upon reasonable request, provided that sharing is consistent with institutional and regulatory requirements.
References
Levine, D. Ultrasound versus magnetic resonance imaging in fetal evaluation. Top. Magn. Reson. Imaging 12, 25–38 (2001).
Aboughalia, H. et al. Multimodality imaging evaluation of fetal spine anomalies with postnatal correlation. Radiographics 41, 2176–2192 (2021).
Prayer, D. et al. ISUOG practice guidelines (updated): performance of fetal magnetic resonance imaging. Ultrasound Obstet. Gynecol. 61, 278–292 (2023).
Parilla, B. V., Leeth, E. A., Kambich, M. P., Chilis, P. & MacGregor, S. N. Antenatal detection of skeletal dysplasias. J. Ultrasound Med. 22, 255–258 (2003).
Bondioni, M. P. et al. Comparative x-ray morphometry of prenatal osteogenesis imperfecta type 2 and thanatophoric dysplasia: a contribution to prenatal differential diagnosis. La Radiologia Medica 122, 880–891 (2017).
Poyner, S. E. et al. Jeune syndrome: Considerations for management of asphyxiating thoracic dystrophy. Neonatal Netw. 32, 342–347 (2013).
Friebe-Hoffmann, U. et al. The femur too short? 1373 fetuses with short femur during second-trimester screening. Arch. Gynecol. Obstet. 306, 1037–1044 (2022).
Papageorghiou, A. et al. Ultrasound-based gestational-age estimation in late pregnancy. Ultrasound Obstet. Gynecol. 48, 719–726 (2016).
D’Ambrosio, V. et al. Midtrimester isolated short femur and perinatal outcomes: a systematic review and meta-analysis. Acta Obstet. Gynecol. Scand. 98, 11–17 (2019).
Gaccioli, F. et al. Femur development in fetal growth restriction as observed on prenatal ultrasound. Ultrasound Obstet. Gynecol. 59, 123–130 (2022).
Tsai, S., Clemente-Casares, X., Revelo, X. S., Winer, S. & Winer, D. A. Are obesity-related insulin resistance and type 2 diabetes autoimmune diseases?. Diabetes 64, 1886–1897. https://doi.org/10.2337/db14-1488 (2015).
Melamed, N. et al. Accuracy of sonographic fetal weight estimation: a matter of presentation. Ultrasound Obstet. Gynecol. 38, 418–424. https://doi.org/10.1002/uog.8928 (2011).
Toijonen, A., Heinonen, S., Gissler, M. & Macharey, G. Risk factors for adverse outcomes in vaginal preterm breech labor. Arch. Gynecol. Obstet. 303, 93–101. https://doi.org/10.1007/s00404-020-05731-y (2021).
Nemec, U., Brugger, P. C., Prayer, D. & Kasprian, G. Skeletal development on fetal magnetic resonance imaging. Top. Magn. Reson. Imaging 22, 131–138 (2011).
Davidson, J. et al. Fetal body MRI and its application to fetal and neonatal treatment. Lancet Child Adolesc. Health The 5, 447–458 (2021).
Nogueira, R. A. et al. The role of a novel magnetic resonance imaging sequence in the evaluation of the fetal skeleton: a pilot study. Radiol. Bras. 51, 303–307 (2018).
Du, J. et al. Qualitative and quantitative ultrashort echo time (UTE) imaging of cortical bone. J. Magn. Reson. 207, 304–311 (2010).
Nemec, S. F. et al. Femur development in fetal growth restriction as observed on prenatal magnetic resonance imaging. Ultrasound Obstet. Gynecol. 61, 601–609 (2023).
Gholipour, A. et al. Fetal MRI: a technical update with educational aspirations. Concepts Magn. Reson. A 43, 237–266 (2014).
Nemec, U. et al. Human long bone development in vivo: Analysis of the distal femoral epimetaphysis on MR images of fetuses. Radiology 267, 570–580 (2013).
Uus, A. Retrospective motion correction in foetal MRI for clinical applications: Existing methods, applications and integration into clinical practice. Br. J. Radiol. https://doi.org/10.1259/bjr.20220071 (2023).
Ponrartana, S. Low-field 0.55 T MRI evaluation of the fetus. Pediatr. Radiol. 53, 1469–1475 (2023).
Aviles Verdera, J. et al. Reliability and feasibility of low-field-strength fetal MRI at 0.55 T during pregnancy. Radiology 309, e223050 (2023).
Matthew, J. et al. A comparison of ultrasound with magnetic resonance imaging in the assessment of fetal biometry and weight in the second trimester of pregnancy: an observer agreement and variability study. Ultrasound 26, 229–244 (2018).
Pietsch, M. et al. APPLAUSE: Automatic prediction of PLAcental health via U-net segmentation and statistical evaluation. Med. Image Anal. 69, 102145. https://doi.org/10.1016/j.media.2021.102145 (2021).
Hoffmann, M. et al. Rapid head-pose detection for automated slice prescription of fetal-brain MRI. Int. J. Imaging Syst. Technol. 31, 1136–1154 (2021).
Silva, Neves & S. Fully automated planning for anatomical fetal brain MRI on 0.55 T. Magnetic Resonance in Medicine (2024) (Epub ahead of print).
Silva, S. N. et al. Automatic flow planning for fetal cardiovascular magnetic resonance imaging. J. Cardiovasc. Magn. Reson. 27, 101888 (2025).
Carion, N. End-to-end object detection with transformers. In European conference on computer vision 213–229 (2020).
Cheng, B., Misra, I., Schwing, A. G., Kirillov, A. & Girdhar, R. Masked-attention mask transformer for universal image segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 1290–1299 (2022).
Strudel, R., Garcia, R., Laptev, I. & Schmid, C. Segmenter: Transformer for semantic segmentation. In Proceedings of the IEEE/CVF international conference on computer vision, 7262–7272 (2021).
Chen, J. et al. TransUNet: Rethinking the U-Net architecture design for medical image segmentation through the lens of transformers. Med. Image Anal. 97, 103280 (2024).
Ronneberger, O., Fischer, P. & Brox, T. U-net: Convolutional networks for biomedical image segmentation. In International Conference on Medical image computing and computer-assisted intervention, 234–241 (Springer, 2015).
Canny, J. A computational approach to edge detection. IEEE Trans. Pattern Anal. Mach. Intell. https://doi.org/10.1109/TPAMI.1986.4767851 (2009).
Cheng, B., Schwing, A. & Kirillov, A. Per-pixel classification is not all you need for semantic segmentation. Adv. Neural Inf. Process. Syst. 34, 17864–17875 (2021).
Cuturi, M. Sinkhorn distances: Lightspeed computation of optimal transport. Adv. Neural Inf. Process. Syst. 26, 2292–2300 (2013).
De Plaen, H. Unbalanced optimal transport: A unified framework for object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition 3198–3207 (2023).
Kirillov, A., Wu, Y., He, K. & Girshick, R. Pointrend: Image segmentation as rendering. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 9799–9808 (2020).
Aviles Verdera, J. et al. Real-time fetal brain and placental t2* mapping at 0.55t mri. Magn. Reson. Med. Sci. 94, 615–624. https://doi.org/10.1002/mrm.30497 (2025).
Chow, K., Kellman, P. & Xue, H. Prototyping image reconstruction and analysis with FIRE. In SCMR 24th Annual Scientific Sessions (2021). Virtual Meeting.
Papageorghiou, A. T. et al. Ultrasound methodology used to construct the fetal growth standards in the INTERGROWTH-21st project. BJOG: An International Journal of Obstetrics & Gynaecology 120, 27–32 (2013).
Namgung, R. & Tsang, R. C. Bone in the pregnant mother and newborn at birth. Clin. Chim. Acta 333, 1–11 (2003).
Kiserud, T. et al. The World Health Organization fetal growth charts: a multinational longitudinal study of ultrasound biometric measurements and estimated fetal weight. PLoS Medicine 14, e1002220 (2017).
Zalevskyi, V. Advances in automated fetal brain MRI segmentation and biometry: insights from the FeTA 2024 challenge. arXiv preprint. arXiv:2505.02784 (2025).
She, J. et al. Automatic biometry of fetal brain MRIs using deep and machine learning techniques. Sci. Rep. 13, 17860 (2023).
Wang, X. et al. Joint segmentation and landmark localization of fetal femur in ultrasound volumes. In 2019 IEEE EMBS International Conference on Biomedical & Health Informatics (BHI), 1–5 (2019).
Slimani, S. et al. Fetal biometry and amniotic fluid volume assessment end-to-end automation using deep learning. Nat. Commun. 14, 7047 (2023).
Shahabodini, S. The missing point in vision transformers for universal image segmentation. arXiv preprint. arXiv:2505.19795 (2025).
Li, F. Dn-detr: Accelerate detr training by introducing query denoising. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition 13619–13627 (2022).
Zareapoor, M., Shamsolmoali, P., Zhou, H., Lu, Y. & García, S. Fractional correspondence framework in detection transformer. In Proceedings of the 32nd ACM International Conference on Multimedia, 5498–5506 (2024).
Kacem, Y. et al. Fetal weight estimation: comparison of two-dimensional US and MR imaging assessments. Radiology 267, 902–910 (2013).
Hadlock, F. P., Harrist, R. B., Sharman, R. S., Deter, R. L. & Park, S. K. Estimation of fetal weight with the use of head, body, and femur measurements–a prospective study. Am. J. Obstet. Gynecol. 151, 333–337 (1985).
Avisdris, N. et al. Automatic linear measurements of the fetal brain on MRI with deep neural networks. Int. J. Comput. Assist. Radiol. Surg. 16, 1481–1492 (2021).
Chen, H., Miller, S., Lane, R. H. & Moyer-Mileur, L. J. Intrauterine growth restriction decreases endochondral ossification and bone strength in female rats. Am. J. Perinatol. 30, 261–266 (2013).
Acknowledgements
The authors thank all pregnant women and their families for participating in this study.
Funding
This work was supported by an MRC project grant [MR/X010007/1], a UKRI FL fellowship [MR/T018119/1] to JH, UKRI DTP funding [2604718] to JAV, DFG Heisenberg funding [502024488] and an ERC StG EARTHWORM [101165242] to JH, and by core funding from the Wellcome/EPSRC Center for Medical Engineering [WT203148/Z/16/Z].
Author information
Authors and Affiliations
Interventional and Experimental Endoscopy (InExEn), Department of Internal Medicine 2, University Hospital Würzburg, Würzburg, 97080, Germany
Johannes Barcsay
School of Biomedical Engineering and Imaging, King’s College London, London, SE1 7EH, United Kingdom
Sara Neves Silva, Mary Rutherford & Jana Hutter
Institute for Information Processing, Leibniz University Hannover, Hannover, 30167, Germany
Jordina Aviles Verdera & Jana Hutter
L3s, Hannover, Germany, 30167
Jordina Aviles Verdera & Jana Hutter
MR Research Collaborations, Siemens Healthcare Limited, Camberley, GU15 3YL, United Kingdom
Raphael Tomi-Tricot
Radiologisches Institut, Universitätsklinikum Erlangen, Erlangen, 91054, Germany
Susanne Schulz-Heise
Authors
- Johannes Barcsay
- Sara Neves Silva
- Jordina Aviles Verdera
- Raphael Tomi-Tricot
- Mary Rutherford
- Susanne Schulz-Heise
- Jana Hutter
Contributions
JB: Writing - Original Draft, Methodology, Software SNS: Methodology, Investigation, Conceptualization JAV: Writing – Review & Editing, Investigation RTT: Writing - Review & Editing MR: Writing - Review & Editing, Funding acquisition SSH: Writing - Review & Editing, Validation JH: Writing - Review & Editing, Supervision, Funding acquisition
Corresponding author
Correspondence to Johannes Barcsay.
Ethics declarations
Competing interests
The authors declare no competing interests.
Additional information
Publisher's note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
Rights and permissions
Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if changes were made. The images or other third party material in this article are included in the article’s Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article’s Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by/4.0/.
About this article
Cite this article
Barcsay, J., Neves Silva, S., Aviles Verdera, J. et al. Real-time AI-guided planning of fetal femur bone MRI and automatic quantification of growth. Sci Rep 16, 31125 (2026). https://doi.org/10.1038/s41598-026-73796-x
Received: 17 April 2026
Accepted: 25 September 2026
Published: 06 October 2026
Version of record: 06 October 2026
DOI: https://doi.org/10.1038/s41598-026-73796-x
来源:Nature 健康护理主题 · nature.com