A Supervised Machine Learning Approach Using Lung Vasculature to Classify Congenital Diaphragmatic Hernia Severity

Introduction: Congenital diaphragmatic hernia (CDH) is a severe birth defect that affects 2–3 in 10,000 births in which the abdominal contents protrude into the thoracic cavity, resulting in lung hypoplasia and altered vascular development [1], [2]. The exact etiology of CDH remains unclear and despite advances in knowledge of CDH and clinical indicators of disease outcome, it remains difficult to predict the disease trajectory following diagnosis with CDH [3]. Current models for studying CDH primarily focus on replicating pulmonary hypoplasia, however vascular changes and pulmonary hypertension (PH) secondary to CDH are major determinants of mortality. Structural changes in the vasculature due to CDH result in reduced pulmonary blood flow, and the severity of this remodeling may serve as a valuable predictor of disease progression [4]. Mice and humans share similar diaphragm and pulmonary development, therefore, we propose studying pulmonary vascular changes using a mouse model of CDH to determine if CDH outcome can be determined from vascular morphology alone. We used a supervised machine learning approach for its ability to recognize complex patterns among data that humans may not immediately recognize, and in this case identify salient regions of the pulmonary vasculature that may be indicative of CDH severity. This will aid in current clinical stratification methods that do not take into account PH for predicting CDH outcomes.

Materials and

Methods: Pregnant CD1 mice were dosed with nitrofen (0.5g/ kg bw) and bisdiamine (0.125 g/kg bw) (N/B) or olive oil vehicle control via oral gavage at embryonic day (E)8.5. Ten days later, fetal lungs were harvested, sectioned, and stained for PECAM-1 and α-smooth muscle actin to visualize the vasculature via fluorescent microscopy. At harvest, fetuses were grouped into 3 groups based on their treatment and resulting phenotype: N/B, +hernia; N/B, -hernia; olive oil control. For the classification algorithm, we used a supervised convolutional neural network modeled after an AlexNet architecture with spatial attention modules following each convolutional layer to classify the images of the fetal lung vasculature by experimental group described above. A windowed-cropping approach was used to subsample whole images into 256x256 pixel-size windows to reduce the computational effort required during training. Our batch size was 8 and we used an 80/20 training/testing split, and 75/25 training/validation split for our dataset. After performing a random search to determine optimal hyperparameters, we selected an initial learning rate of 0.0001 scheduled to decrease by a factor of 0.05 every 10 epochs. Loss was calculated using unweighted cross-entropy loss criterion, and an Adam optimizer was used to update parameters in a direction that minimizes the loss, training to a total of 50 epochs.

Results, Conclusions, and Discussions: Following training and validation, our model framework had an overall testing classification accuracy of 65%. Due to the complexity of the experimental images classified relative to the training dataset, we used a smaller batch size with a larger number of epochs to train our model. Following testing, our model more correctly classified images from control animals and images of animals dosed with N/B with a hernia, compared to images from samples that were dosed with N/B and did not have a hernia (Figure 1A). This indicated that more significant morphometric changes from the control may occur in the vasculature between animals that did develop a hernia compared to those that did not after N/B dose, which we expected based on the observable differences between classes with a hernia and without. We then randomly shuffled the labels in the testing dataset and evaluated the accuracy of our trained model. This random shuffling resulted in an accuracy of 35%, a decrease from our initial accuracy of 65%. Our trained model classified instances of CDH better than random, showing our model’s potential to serve as a tool for predicting CDH outcome based on the pulmonary vasculature alone.
Attention maps were printed to visualize regions within images that the model used to differentiate between classes (Figure 1BC). The attention maps did not highlight any noticeable features of stained vasculature, although our model did appear to focus attention on regions where the density of the vasculature appeared higher. This lack of a visible pattern found to classify our images may also indicate that our model recognized a pattern too complex for visualization using this strategy. Whereas our model could not select visible regions within the vasculature that indicate CDH severity, it was able to classify CDH outcome based on vasculature alone, highlighting that vascular remodeling changes may be a relevant indicator of CDH severity. Outcomes of this study demonstrate the potential for deep learning and visualization of pulmonary vasculature to inform clinical stratification for CDH severity.

Acknowledgements (Optional):

References: [1] Rivas, J.F.G. et al, Pediatr. Res. 95, 912–921, 2024. [2] Dumpa, V. et al, StatPearls, 2023. [3] Daodu, O. et al, Seminars in Pediatric Surgery, 2017. [4] Patel N. et al, Front Pediatr. 2022.