A probabilistic deep learning approach for choroid plexus segmentation in autism spectrum disorder

Area of Science:

• Neuroimaging
• Artificial Intelligence in Medicine
• Developmental Neuroscience

Background:

• The choroid plexus is a critical brain barrier involved in neuroimmune function.
• Morphological alterations in the choroid plexus are observed in some individuals with autism spectrum disorder (ASD).
• Accurate segmentation of the choroid plexus using magnetic resonance imaging (MRI) is essential for large-scale ASD population studies.

Purpose of the Study:

• To evaluate the generalizability of the Automated Segmentation of CHOroid PLEXus (ASCHOPLEX) deep learning tool for choroid plexus segmentation in individuals with ASD.
• To assess the performance of ASCHOPLEX across different age groups (children and adults).
• To investigate the utility of a probabilistic approach for quantifying segmentation uncertainty and model confidence.

Main Methods:

• Finetuning of the ASCHOPLEX deep learning tool on a local dataset of ASD and control participants.
• Implementation of a probabilistic version of ASCHOPLEX to quantify segmentation uncertainty.
• Testing ASCHOPLEX performance on the Autism Brain Imaging Data Exchange (ABIDE) dataset, comprising both children and adults.

Main Results:

• ASCHOPLEX demonstrated good generalizability and accurate choroid plexus segmentation in adults across both local and ABIDE datasets.
• The tool's accuracy decreased in children, indicating limited generalizability to this age group without further finetuning.
• The probabilistic approach provided confidence metrics, enhancing the reliability assessment of the segmentation model.

Conclusions:

• ASCHOPLEX is a capable deep learning tool for choroid plexus segmentation in previously unseen MRI data, particularly in adults.
• Age-specific finetuning is necessary to ensure ASCHOPLEX's accuracy and generalizability in pediatric populations.
• Incorporating probabilistic methods strengthens the utility of deep learning tools by providing crucial confidence assessments for clinical and research applications.