Improving Prenatal Alcohol Exposure classification using Data Augmentation in 3D Convolutional Neural Networks

Prenatal alcohol exposure (PAE) is the consumption of alcohol by a mother during pregnancy, potentially leading to a range of developmental and behavioral problems in the child. Early detection is vital to mitigate these problems. In addition to clinical assessment, magnetic resonance (MR) scans offers a non-invasive method to examine the brain's detailed structure, enabling researchers to explore subtle anatomical variations associated with PAE. Machine learning, particularly <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$3\mathrm{D}$</tex> Convolutional Neural Networks (CNNs), serves as a powerful tool to automatically identify patterns in the children's brains. However, the limited size of available PAE data creates a significant bottleneck for developing effective models. To address this, our goal is to explore data augmentation techniques to artificially expand the dataset, thereby improving model performance and generalizability. We assessed three rates of data augmentation <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$(1x,\ 3x,\ 6x)$</tex> in six 3D CNN models with increasing architectural complexity. Adding data augmentation significantly improved validation accuracy <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$(3x$</tex> and <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$6x \quad vs \quad1x)$</tex>, but no significant difference was found between <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$3x$</tex> and <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$6x$</tex>. Models with fewer layers and balanced number of parameters had higher accuracy. In conclusion, simpler CNN architectures generalized better, and data augmentation aids model convergence, though benefits plateau beyond <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$3x$</tex> augmentation. Future work includes adding interpretability modules to identify brain differences among participants.