.png){kind=small}
MIND
Mental ILLNESS rECOGNITION
THROUGH NEUROIMAGING AND
DEEP lEARNING
The Problem
The physiological conditions underlying psychiatric disorders are not well understood at present. Current psychiatric diagnoses rely on clinical observations and symptomatic criteria, leading to a lack of clear boundaries between different disorders, as well as high heterogeneity within the same disorder category.
Our Goal
In the future, the hope is to have more robust diagnostic criteria, potentially based upon neuroscience and other biomarkers instead of behavioral characteristics. Eventually, this may lead to the development of additional diagnostic tests using a patient’s biological data.
Before that can be done however, the connection between biological systems and expressed symptoms must be established.
​
Our research question: do neuronal biomarkers correspond to current classifications of mental illness?
Project Approach
-
Develop Supervised and Unsupervised models
-
Use the models to visualize the data using clustering
-
Use the models to identify important areas of the brain
heterogeneity & Symptom overlap
In research and medicine, patient and control populations are assumed to be well-defined and consistent within their groups. This has the potential to be particularly problematic in psychiatry, as disorders are categorized based on symptoms, and issues of over- and under-specificity arise.
One result of this is a significant amount of overlapping symptoms between different disorders, suggesting that the current boundaries are not as clear as they could be [1]. Another result is a high degree of variability within disorders, termed the “heterogeneity problem.” This refers to the existence of numerous disease subtypes, as well as the exceedingly large variety of potential symptom combinations that makes it possible for two individuals to receive the same diagnosis while sharing none of the same symptoms [1,2,3].
A) Conventional assumption that patients and controls form uniform and distinct groups.
B) Patients form multiple groups with distinct pathology.
C) Disease-related variation may be nested within healthy variation.
D) Patients may be diffuse and heterogeneous due to misdiagnosis, comorbidities, or an aggregation of different pathologies.
Supervised Model
Goals:
-
Find patterns in the clustering of current labeled types.
-
Overlaps indicate issues with heterogeneity/ symptom overlap since they are present as the same biologically
Approach:
-
3D MRI scans of patients with a corresponding disease label
-
Train deep convolutional model with voxresnet [4] architecture
-
Implemented using deformable 3D convolutions [5]
