Radiomics have the potential to enhance the assessment of fetal MRI data by extracting quantitative image features that may not be visually perceivable [20]. However, reliable radiomics-assisted fetal-MRI-based tissue characterisation requires excellent feature reproducibility [11]. The presented findings show that the use of 2D versus 3D lung ROIs for radiomics feature extraction from fetal MRI data severely impacts feature values. In addition, radiomics features extracted from 3D ROIs encompassing the whole fetal lung outperformed features extracted from 2D lung ROIs with regard to reproducibility in repeated image acquisitions. Therefore, in the future, highly-reproducible fetal MRI radiomics features extracted from whole lung segmentation masks may improve non-invasive quantitative assessment of lung development.

Non-invasive in-vivo MRI assessment of the fetal lungs is safe and feasible during the second and third trimesters, which corresponds to the canalicular and saccular phases of lung development [1]. During this time, besides volume growth, lung organogenesis is characterised by microstructural changes including the formation of pulmonary acini, differentiation of type I and II pneumocytes, and increasing production of lung fluid and surfactant. This can be observed in fetal MRI as an increase in lung volume along with an increase in signal intensity [21]. However, visual assessment of fetal lung signal intensity in MRI is subjective. Integration of quantitative measures of tissue characteristics in the form of lung-to-liver, lung-to-muscle, or lung-to spinal fluid signal intensity ratios into fetal MRI lung assessment have produced mixed results [4, 22,23,24], so far prohibiting their translation into clinical routine. Therefore, current image-based assessment of fetal lung development focuses primarily on tissue quantity in the form of lung volume rather than tissue quality. Unfortunately, volume alone is an imperfect descriptor of lung developmental status as gestational age-adjusted growth curves show wide normal ranges [2]. Recently, the use of novel fetal MRI techniques including diffusion-weighted imaging [25], intra-voxel incoherent motion analysis [26] and T2* mapping [27] for the microstructural characterisation of fetal lung tissue has been advocated but their benefit remains largely unclear.

In order to facilitate non-invasive, image-based and timely detection of abnormal fetal lung development, reliable quantitative lung tissue features beyond volume are needed. Radiomics allows the extraction of a multitude of features reflecting various aspects of shape and texture from 2D or 3D image ROIs [28]. Fetal MRI is ideally suited for the extraction of quantitative lung radiomics features since image acquisition follows a standardised protocol. Most fetal imaging centres use a single MRI scanner, which has been shown to be essential for radiomics feature reproducibility. Moreover, dedicated fetal MRI lung assessments already requires manual whole lung segmentation to obtain lung volume [29]. These lung segmentations could be integrated into a post-processing pipeline for radiomics feature extraction that has potential for automation due to the high level of standardisation recommended by the Image Biomarker Standardisation Initiative [30]. Therefore, fetal MRI radiomics analysis of lung development could be implemented without the need for additional costly resources.

Fetal lung texture analysis has thus far only been explored using ultrasound images: Previous studies assessed lung maturity [31], identified cases at risk for pulmonary hypoplasia [32], and predicted neonatal respiratory insufficiency [33] with promising results. However, ultrasound studies used 2D fetal lung ROIs to extract texture features, e.g. from lung tissue at the level of the four chamber view. In addition, physicians placed ultrasound ROIs in ‘representative lung areas’ while avoiding artifacts. This approach raises questions with regard to the reliability of radiomics-based fetal lung assessment: Cardiac position (and the level of the four chamber view) may vary in pathologies where lung development is of particular interest, such as in fetuses with congenital diaphragmatic hernia, or congenital pulmonary airway malformation. Furthermore, it is unclear whether a two-dimensional ROI sufficiently represents tissue characteristics of a much larger three-dimensional structure, particularly if subjective criteria are used to determine the ROI’s location. Lastly, there is a lack of evidence regarding the reproducibility of radiomics features extracted from 2D ROIs in repeated image data acquisitions.

The presented results suggest that radiomics features extracted from 2D fetal MRI lung ROIs do not adequately represent whole lung tissue characteristics. Few features (5 of 95 [5.3%]; 10Percentile, Mean, RootMeanSquared, Median, 90Percentile) showed excellent reproducibility between the use of 2D and 3D ROIs in fetal MRI data from a single acquisition. Moreover, these features all reflected basic first order statistics but none shape or higher-order texture, effectively excluding potentially crucial information contained within the lung’s microstructure from further analysis. Another essential finding concerns the limited radiomics feature reproducibility for the use of 2D lung ROIs in repeated fetal MRI acquisitions. Here, excellent reproducibility was found in only twelve of 95 (12.6%) features. Notably, this was achieved using standardised 2D ROIs that covered the entire lung tissue except hilar structures on an axial fetal MRI slice at the level of the carina. Subjective segmentation of representative-appearing lung areas may have further increased radiomics feature variability. In contrast, excellent radiomics feature reproducibility was found in a majority of features (49 of 95, 51.6%), including a variety of shape and higher-order texture features, if 3D lung ROIs were used for both fetal MRI acquisitions.

The complex and dynamically changing microstructure of the fetal lung are unlikely to be adequately reflected by one quantitative parameter, which may explain the limited utility of previously explored lung signal intensity ratios for outcome prediction [23, 24]. Rather, a combination of images markers, i.e. ‘a radiomics signature’ focusing on different fetal lung characteristics, such as texture and shape, may be more suitable as a quantitative descriptor of lung development. Undoubtedly, lung volume remains an essential parameter to determine whether lung development is age-appropriate. By adding texture and shape features to a lung radiomics signature, the sensitivity of fetal MRI for the detection of abnormal lung development may be increased, e.g. in cases where lung volume is within normal ranges but lung microstructure is altered. As demonstrated, whole lung fetal MRI radiomics enables the extraction of a large number of highly-reproducible lung shape and texture features that may be used to develop radiomics signatures of normal and pathologic lung development in the future.

As a proof of concept, we tested whether 3D and 2D ROIs were sufficient to detect significant differences between radiomics features extracted from fetal lungs with normal or pathological development in our limited sample size. We found 11% of radiomics features were significantly different between lungs with normal compared to pathological development when using 3D ROIs. Critically, in case of the use of 2D ROIs, no significant difference in radiomics features was observed, indicating that radiomics features extracted from 3D ROIs are more sensitive to subtle, visually not perceivable changes in the fetal lung’s microstructure. Further studies are necessary to confirm these findings and identify robust and predictive 3D fetal lung radiomics features.

This study had several limitations: The number of included cases was relatively small but similar or larger compared to previous test–retest studies on radiomics feature reproducibility [13,14,15]. A single 1.5 T MRI scanner was used. Due to the retrospective study design, routinely performed repeated MRI acquisitions from a single examination were utilised rather than repeated fetal MRI scans. While T2-weighted images are widely used for visual assessment of the fetal lung [34], it is not known which MRI sequence is best suited for radiomics-based analysis. The radiomics feature set included in the current analysis is limited, but is compatible with the Image Biomarker Standardisation Initiative guidelines [30]. Pyradiomics has been widely used in lung imaging and beyond facilitating comparability and generalizability of results [18]. Excellent radiomics feature reproducibility was conservatively defined as ICC > 0.9, but there is a lack of evidence concerning an optimal cut-off value.