Disentangled Hybrid Transformer for Identification of Infants with Prenatal Drug Exposure.

Prenatal drug exposure, which occurs during a time of extraordinary and critical brain development, is typically associated with cognitive, behavioral, and physiological deficits during infancy, childhood, and adolescence. Early identifying infants with prenatal drug exposures and associated biomarkers using neuroimages can help inform earlier, more effective, and personalized interventions to greatly improve later cognitive outcomes. To this end, we propose a novel deep learning model called disentangled hybrid volume-surface transformer for identifying individual infants with prenatal drug exposures. Specifically, we design two distinct branches, a volumetric network for learning non-cortical features in 3D image space, and a surface network for learning features on the highly convoluted cortical surface manifold. To better capture long-range dependency and generate highly discriminative representations, image and surface transformers are respectively employed for the volume and surface branches. Then, a disentanglement strategy is further proposed to separate the representations from two branches into complementary variables and common variables, thus removing redundant information and boosting expressive capability. After that, the disentangled representations are concatenated to a classifier to determine if there is an existence of prenatal drug exposures. We have validated our method on 210 infant MRI scans and demonstrated its superior performance, compared to ablated models and state-of-the-art methods.