Unsupervised artificial intelligence models for Down syndrome neuroimaging research: from automatic brain alteration detection to the generation of synthetic 3D magnetic

AI Technology: Deep Learning; Vision (image recognition, image generation, text recognition OCR, etc.) and other. 

Brain imaging techniques, particularly magnetic resonance imaging (MRI), play a crucial role in understanding the neurocognitive phenotype of Down syndrome (DS) and its associated challenges.

MRI provides detailed insights into the structural and functional alterations in the brain due to genetic imbalances caused by trisomy 21. Individuals with DS exhibit distinctive craniofacial features, like microcephaly and brachycephaly, alongside reduced brain volume, especially in the hippocampus and cerebellum. Structural anomalies like ventriculomegaly and corpus callosum malformations are also common, contributing to cognitive impairments and early-onset dementia.

Despite advancements, the links between cognitive performance and brain anatomy, as well as neuroinflammation and comorbidities like Alzheimer’s disease, remain unclear. 

The complexity of analyzing brain MRI scans requires expertise and time, prompting the exploration of artificial intelligence (AI) for automated assistance. Unsupervised AI techniques, particularly Autoencoders (AE), offer a solution by learning the distribution of healthy brain anatomy and detecting alterations in unseen scans. 

AE compress MRI scans into a latent space during training, enabling the detection of pathological features during inference. Moreover, Latent Diffusion Models (LDM) facilitate the generation of synthetic brain MRI scans, addressing data scarcity and privacy concerns.In this project, we aim to develop unsupervised AI techniques for assessing brain alterations in DS. 

The project team plans to train AE with both real and synthetic 3D brain MRI scans to enhance detection accuracy. Additionally, through LDM we will generate synthetic brain scans representing non-pathological DS-associated features, addressing the challenge of obtaining diverse datasets without privacy concerns. 

These synthetic datasets will enable the training of AE models tailored to DS, allowing investigation into associated comorbidities like Alzheimer’s disease.