Triplet-Based Deep Hashing Incremental Learning for Brain Network Classification.

Area of Science:

• Neuroscience
• Machine Learning
• Data Science

Background:

• Public brain network datasets often combine data from multiple sites, leading to performance issues due to data heterogeneity.
• Variations in data collection across sites negatively impact the accuracy of brain network classification models.

Purpose of the Study:

• To introduce a novel Triplet-based Deep Hashing Incremental Learning (Tri-DHIL) method for robust brain network classification.
• To address the challenge of multi-source data heterogeneity in brain imaging datasets by enabling incremental learning from individual sites.

Main Methods:

• The Tri-DHIL method involves three phases: site queue generation, triplet-based deep hashing learning, and incremental learning.
• Site ranking is based on sample quantity and label information. Samples are clustered using diagnostic labels to form triplets (anchor, same cluster, different cluster).
• Deep hashing learning extracts features and maps them to hash codes. Incremental learning adjusts model parameters using accumulated triplet-based losses to prevent catastrophic forgetting.

Main Results:

• Experimental results on ABIDE I, ABIDE II, and ADHD-200 datasets demonstrate the effectiveness of the Tri-DHIL method.
• The proposed method achieves competitive classification performance despite data heterogeneity from multiple sites.
• Incremental learning successfully prevents the model from forgetting previously learned features when incorporating new site data.

Conclusions:

• The Tri-DHIL method offers a promising solution for brain network classification with multi-site datasets.
• Incremental learning is crucial for adapting models to heterogeneous data streams without compromising previously acquired knowledge.
• This approach enhances the practical applicability of machine learning models in neuroimaging research by handling real-world data complexities.