Introduction: Autism spectrum disorder (ASD) is a developmental disorder, affecting about 1% of the global population. It is characterized by problems with impaired linguistic, communication, cognitive and social skills. Symptoms of ASD vary in severity between individuals. Therapies exist that treat the symptoms, provide information about the condition, and improve patient quality of life. However, the diagnosis of autism presents a challenge because a basic medical test has not yet been developed. To date, behavioral tests remain the only approach to diagnosing ASD. This, however, requires prolonged diagnostic time and increased medical costs.
Early and accurate diagnosis and prognosis are urgently needed for this neurological disorder. Integration of resting-state functional magnetic resonance imaging (rs-fMRI) techniques and machine learning algorithms are showing great promise in unveiling hidden pathological functional connectome patterns to assist early diagnosis and prediction of brain disorders. However, brain functional connectome patterns analysis remains challenging due to the inherent high dimensionality of data and insufficient sample sizes.
Much of human learning involves only a few new examples superimposed on extensive prior knowledge. Motivated by how human learn new knowledge, transfer learning focuses on storing knowledge gained from solving problems in one data-rich domain and applying it to a new problem in another data-scarce domain. Transfer learning represents an important key to solve the fundamental problem of insufficient training data in deep learning.
Our group proposed to pretrain a deep learning model prototype by utilizing relatively easy-to-obtain functional connectome from a database of healthy subjects and then transferred this prototype to build a fine-tuned model for the ASD diagnosis task that had limited training samples. (Li, Parikh et al. 2018) We demonstrated that our model achieved enhanced ASD classification as compared to peer models without the transfer learning strategy. The significantly improved performance was observed irrespective of site sample size and was reproducible among various subsampling schemes. The proposed model identified the functional connection between the left superior occipital gyrus and right inferior occipital gyrus as the most significant feature to distinguish ASD patients.
Li H, Parikh NA, & He L. (2018). A Novel Transfer Learning Approach to Enhance Deep Neural Network Classification of Brain Functional Connectomes. Front Neurosci, 12, 491. PMID: 30087587; PMCID: PMC6066582. Pubmed Journal
Parikh MN, Li H, He L. (2019). Enhancing Diagnosis of Autism with Optimized Machine Learning Models and Personal Characteristic Data. Front Comput Neurosci. 15, 13-9. PMID: 30828295; PMCID: PMC6384273. Pubmed Journal