Objectives
The Carnegie staging system facilitates the assessment of normal and abnormal development in terms of morphology during the embryonic period. Using virtual reality (VR) it is possible to visually assess the Carnegie stage in-utero, which takes 1-2 minutes per ultrasound image. Adoption in clinical practice is hampered by the need for a VR set-up and required time for visual assessment. To overcome this, our aim is to automate in-utero Carnegie staging using Artificial Intelligence (AI).

Methods
1357 first trimester three-dimensional (3D) ultrasound images of 797 ongoing pregnancies resulting in life birth from The Rotterdam Periconception Cohort were used. We used DenseNet, a state-of-the-art deep learning algorithm for image classification. The algorithm was trained to estimate the Carnegie stage < 16, 16-23, and 23> solely based on the ultrasound images. We used 1100 images of 642 pregnancies for training. For evaluation, we used a test set of 257 images of 155 pregnancies, not used during training.

Results
The AI algorithm achieved an overall accuracy of 61%, which is close to the results of an independent rater, who achieved an accuracy of 63% on 46 images selected for manual VR assessment training. The accuracy was for stage < 16: 55% (n = 9), for stages 16-19: 59% (n = 79), for stages 20-23:62% (n = 151), and for stage >23: 61% (n = 18). The performance differences can partly be explained by the limited size of the embryo early in the first trimester.

Conclusions
Since automatic Carnegie staging using AI is real-time and does not require a VR set-up adoption in clinical practice becomes feasible. In future work, we aim to enhance interpretability by analysing the specific morphological aspects in ultrasound scans utilised by the algorithm to assign the Carnegie stage. Understanding the morphological aspects linked to the Carnegie stage by the algorithm might lead to more in-depth insight into the patterns of normal and abnormal morphological development across pregnancies.

In this paper, we propose the use of Recurrent Inference Machines (RIMs) to perform T₁ and T₂ mapping. The RIM is a neural network framework that learns an iterative inference process based on the signal model, similar to conventional statistical methods for quantitative MRI (QMRI), such as the Maximum Likelihood Estimator (MLE). This framework combines the advantages of both data-driven and model-based methods, and, we hypothesize, is a promising tool for QMRI. Previously, RIMs were used to solve linear inverse reconstruction problems. Here, we show that they can also be used to optimize non-linear problems and estimate relaxometry maps with high precision and accuracy. The developed RIM framework is evaluated in terms of accuracy and precision and compared to an MLE method and an implementation of the Residual Neural Network (ResNet). The results show that the RIM improves the quality of estimates compared to the other techniques in Monte Carlo experiments with simulated data, test-retest analysis of a system phantom, and in-vivo scans. Additionally, inference with the RIM is 150 times faster than the MLE, and robustness to (slight) variations of scanning parameters is demonstrated. Hence, the RIM is a promising and flexible method for QMRI. Coupled with an open-source training data generation tool, it presents a compelling alternative to previous methods.

For the segmentation of magnetic resonance brain images into anatomical regions, numerous fully automated methods have been proposed and compared to reference segmentations obtained manually. However, systematic differences might exist between the resulting segmentations, depending on the segmentation method and underlying brain atlas. This potentially results in sensitivity differences to disease and can further complicate the comparison of individual patients to normative data. In this study, we aim to answer two research questions: 1) to what extent are methods interchangeable, as long as the same method is being used for computing normative volume distributions and patient-specific volumes? and 2) can different methods be used for computing normative volume distributions and assessing patient-specific volumes? To answer these questions, we compared volumes of six brain regions calculated by five state-of-the-art segmentation methods: Erasmus MC (EMC), FreeSurfer (FS), geodesic information flows (GIF), multi-atlas label propagation with expectation–maximization (MALP-EM), and model-based brain segmentation (MBS). We applied the methods on 988 non-demented (ND) subjects and computed the correlation (PCC-v) and absolute agreement (ICC-v) on the volumes. For most regions, the PCC-v was good ((Formula presented.)), indicating that volume differences between methods in ND subjects are mainly due to systematic differences. The ICC-v was generally lower, especially for the smaller regions, indicating that it is essential that the same method is used to generate normative and patient data. To evaluate the impact on single-subject analysis, we also applied the methods to 42 patients with Alzheimer’s disease (AD). In the case where the normative distributions and the patient-specific volumes were calculated by the same method, the patient’s distance to the normative distribution was assessed with the z-score. We determined the diagnostic value of this z-score, which showed to be consistent across methods. The absolute agreement on the AD patients’ z-scores was high for regions of thalamus and putamen. This is encouraging as it indicates that the studied methods are interchangeable for these regions. For regions such as the hippocampus, amygdala, caudate nucleus and accumbens, and globus pallidus, not all method combinations showed a high ICC-z. Whether two methods are indeed interchangeable should be confirmed for the specific application and dataset of interest.