Fig 1.
Standardized Protocol ItemsRecommendations for Observational Retrospective Study (SPIROS) flow diagram: Schedule for administrative and ethics clearance; research training; retrospective data collection; definition of instrumental parameters for the identification of prognostic patterns; definition of protocols for data management; software system design development of the method for the management of patterns (and data augmentation) in training; development of the method for optimizing system parameters; fetal MRI data elaboration with manual segmentation; development of semiautomatic segmentation software; segmentation performance analysis; fetal MRI data elaboration with semiautomatic segmentation; development of the classification system based on machine learning; development of the classification system based on deep learning; statistical analysis for the coupling of classification methods; statistic and classification analysis; completion of analyses; manuscript writing; result dissemination; kick-off and project meetings. Unit 1: Milan; Unit 2: Lecce; Unit 3: Palermo.
Fig 2.
CDH: Congenital diaphragmatic hernia; GA: Gestational age; NICU: Neonatal intensive care unit; MRI: Magnetic resonance imaging; PH: Pulmonary hypertension; FETO: Fetal endoscopic tracheal occlusion; ECMO: Extracorporeal membrane oxygenation.
Fig 3.
Total fetal lung volume assessment.
MRI segmentation and 3D reconstruction of the fetal lung with the calculation of the total fetal lung volume on T2-sequences.
Fig 4.
MRI segmentation and 3D reconstruction of the fetal liver with the calculation of the herniated liver on T2-sequences.
Fig 5.
MRI calculation of the mediastinal shift angle (MSA) on axial TrueFisp-sequences.
Fig 6.
Apparent diffusion coefficient.
MRI calculation of the apparent diffusion coefficient (ADC) on diffusion-weighted sequences.
Fig 7.
Chest radiographic pulmonary area.
Calculation of the radiographic pulmonary area on neonatal chest x-ray.
Fig 8.
The flowchart illustrates the radiomics workflow starting from multimodal image acquisition. After manual or (semi)automatic segmentation/contouring of the volumes of interest (such as fetus lungs and liver), feature descriptors are calculated from VOI shape and texture. Together with a choice of prenatal clinical parameters, the obtained feature vectors are labeled with relevant output variables (e.g., presence of postnatal PH, or need for ECMO…) and, after dimensionality reduction performed to get rid of redundant and useless descriptors, enter the supervised classification/regression model-construction step, here represented by one of the possible choices, i.e., a multi-layer perceptron artificial neural network. Because of the relatively low number of samples, model training and validation will be obtained by leave-one-out cross-validation (LOO-CV), and various methods, such as the ROC curve, will quantify model quality. The trained and validated ML model will be the pipeline output to be employed in the Decision Support System.
Fig 9.
Schematic of the architecture of a convolutional neural network (CNN).
A CNN is composed of several kinds of layers, namely the convolution layers and the pooling layers. One of the most significant differences between deep networks and other ML algorithms is the use of ReLU as a transfer function to make the algorithm faster. Then, the outputs generated by the previous levels are "flattened" to transform them into a single vector that can be used as an input for the next level. The fully connected layer applies weights to the input generated by the feature analysis to predict an accurate label. Finally, the fully connected output layer and softmax produce the final outputs in order to determine the class to associate with the image.