Transfer Learning for Improving Neuroimaging-Based Diagnostic Classification.

Area of Science:

• Neuroimaging
• Machine Learning
• Artificial Intelligence

Background:

• Overfitting in neuroimaging machine learning hinders diagnostic classification accuracy due to small clinical sample sizes.
• Acquiring large clinical datasets is challenging and expensive, leading to smaller sample sizes compared to healthy controls.
• Existing data aggregation efforts like the Autism Brain Imaging Data Exchange (ABIDE) still face limitations with sample size.

Purpose of the Study:

• To address overfitting in autism diagnostic classification by utilizing larger healthy control datasets.
• To transfer knowledge from healthy brain signatures to improve the discrimination of autism from controls.
• To enhance the generalizability and accuracy of machine learning models in neuroimaging.

Main Methods:

• Developed a variational autoencoder-based transfer learning framework.
• Incorporated data oversampling, model pre-training, and classifier training and testing.
• Estimated and visualized the performance of the transfer learning approach.

Main Results:

• The transfer learning model achieved approximately 7% higher accuracy on site-mismatched data compared to models without transfer learning.
• Demonstrated improved diagnostic classification performance by leveraging larger healthy control datasets.

Conclusions:

• Transfer learning is applicable within a deep learning framework to improve autism diagnosis.
• Utilizing larger healthy control datasets enhances generalizability and accuracy while reducing inter-site variability.
• The proposed framework shows potential for application in diagnosing other neurological and psychiatric disorders.