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

Abstract

The choroid plexus serves as the primary barrier between the brain's blood and cerebrospinal fluid and mediates neuroimmune function. A subset of individuals with autism spectrum disorder (ASD) may exhibit morphological alterations of the choroid plexus. However, to power larger population analyses, an automated tool capable of accurately segmenting the choroid plexus based on magnetic resonance imaging (MRI) is needed. Automated Segmentation of CHOroid PLEXus (ASCHOPLEX) is a deep learning tool that enables finetuning using new, patient-specific, training data, allowing its usage across cohorts for which the model was not originally trained. We evaluated ASCHOPLEX's generalizability to individuals with ASD by performing finetuning on a local dataset of ASD and control (CON) participants. To assess generalizability, we implemented a probabilistic version of the algorithm, which allowed us to quantify the uncertainty in choroid plexus segmentation and evaluate the model's confidence. ASCHOPLEX generalized well to our local dataset, in which all participants were adults. To further assess its performance, we tested the algorithm on the Autism Brain Imaging Data Exchange (ABIDE) dataset, which includes data from children and adults. While ASCHOPLEX performed well in adults, its accuracy declined in children, suggesting limited generalizability to different age groups without additional finetuning. Our findings show that the incorporation of a probabilistic approach during finetuning can strengthen the use of this deep learning tool by providing confidence metrics which allow assessing model reliability. Overall, our findings demonstrate that ASCHOPLEX can generate accurate choroid plexus segmentations in previously unseen data.

Fig. 1. Pipeline for the finetuning of ASCHOPLEX for the deterministic (top) and probabilistic (bottom) variants.

The probabilistic model is obtained by activating the dropout layers within the model ensemble prior to finetuning. Subsequently, the pipeline is identical for both models: a local dataset (n = 65) is split into two groups, with one group of 12 subjects used for finetuning and one group of 53 subjects used for model evaluation.

Fig. 2. Pipeline for the model evaluation of the Finetuned (top) and Probabilistic Finetuned (bottom) version of the ASCHOPLEX algorithm.

To obtain the binary segmentation masks from the probabilistic model we thresholded the output mean prediction probabilities at 0.5 (50%). Subsequently, the pipeline was identical for both models: the binary masks were compared with the available manual segmentations to compute the Dice coefficient.

Fig. 3. Example of the choroid plexus segmentation for an individual subject from the local dataset for each of the methods.

The subject shown is the subject with the best Dice coefficient for ASCHOPLEX Finetuned Probabilistic (A). Dice coefficients for the automated segmentation methods vs. manual segmentation in the 53 held-out participants from the local dataset (B). Dice coefficients compared across sex and diagnosis for each of the segmentation methods for participants from the local dataset (C).

Fig. 4. Model uncertainty computed for the different groups of subjects (adults from local dataset, adults from ABIDE and children from ABIDE).

We computed four uncertainty metrics: (i) Standard deviation of the predictions, (ii) total predicted entropy (H), (iii) aleatoric uncertainty proxy (expected per sample entropy), and (iv) epistemic uncertainty (mutual information). ASD: Autism Spectrum Disorder, CON: Controls, Adults: 18 years old or older, Children: younger than 18 years old.