Using machine learning for classifying and analyzing microscopy sections of brain areas in autism spectrum disorder
Computational Neuroscience & Vision Lab
Supervisor: Prof. Arash Yazdanbakhsh (Boston University)
May 2021-Present
Abstract:
Neuroanatomists are able to identify axon pattern markers in brains from a healthy control population (CTR) versus patients with Autism Spectrum Disorder (ASD); yet, this process demands labor intensive surveys of microscopic sections. We propose a machine learning method to automatically classify microscopic sections from ASD and CTR brains, also taking into account the different white matter regions: superficial white matter (SWM) and deep white matter (DWM). The result is a deep neural network that has successfully classified between ASD and CTR for up to 95% accurate, while still encountering some trouble with DWM and SWM classification (average accuracy of 78%). 3D plotting and sensitivity maps further prove white matter origin to be the base of the network classification confusion, while reassuring the consistency of ASD vs CTR classification with many shared anatomical markers. Utilizing image pre-processing methods also show promising result, especially in aiding with DWM and SWM identification. With more training, this method could make this labor-consuming process more automated, with the potential to discover other distinguishing anatomical markers among ASD patients that could have been overlooked by the human eyes.
Four classes of classification
Autism Spectrum Disorder Deep white matter (ASD DWM)
Autism Spectrum Disorder Superficial white matter (ASD SWM)
Typically Developed Deep white matter (CTR DWM)
Typically Developed Superficial white matter(CTR SWM)
Methods
Convolution Neuro Network
Figure 1a.
Figure 1b.
Figure 1c.
The result of the network classification is then visualized through a series of confusion matrices. The overall accuracy rate for classification over 4 classes is displayed on Figure 1a. To distinguish the network accuracy between ASD versus CTR microscopic sections, we compounded the confusion matrix within two groups (Figure 1b).We combined ASD DWM and ASD SWM as one whole category labeled ASD, and combined CTR DWM and CTR SWM as the other category labeled CTR (Figure 1b). The confusion matrix in Figure 1b demonstrates the effectiveness of the network in classifying axon patterns of CTR (control) and ASD brains. Similarly, the same method was applied to create the confusion matrix in Figure 1c.