Ultrasound technology employs sound waves to generate images of internal body structures, and fetal ultrasound serves as a secure, non-invasive method to observe and assess particular regions of a developing fetus. Digital image processing (DIP) techniques can enhance these images, helping specialists to identify conditions such as congenital diaphragmatic hernia (CDH), Congenital Cystic Adenomatoid Malformation (CCAM) and bronchopulmonary sequestration (BPS), all of which can lead to respiratory distress syndrome (RDS). A significant challenge in this field is the lack of preserved medical sonograms for training robust diagnostic models. Our study addressed this issue by using a dataset from public online repositories. However, due to variations in acquisition protocols, equipment, artefacts, and annotations, these images were highly heterogeneous. To ensure data consistency and mitigate this variability, we implemented extensive preprocessing and data augmentation. We used a Vision Transformer (ViT) model with a ViT16 architecture, which is effective for processing limited and heterogeneous medical imaging datasets. After training the model on the preprocessed fetal lung ultrasound images, we achieved impressive results: 93.75% accuracy and 99% area under the ROC curve (AUC). These metrics highlight the model’s strong ability to accurately classify various fetal lung pathologies despite data limitations and variations. This work demonstrates the potential of ViT models as a powerful tool for enhancing the accuracy of prenatal diagnoses.