Interpretable ROI Identification in Brain Image Analysis: Overcoming CNN Black Box Challenges With Kriging-Enhanced...

Area of Science:

• Neuroimaging
• Computational Neuroscience
• Medical Image Analysis

Background:

• Brain image analysis faces challenges in precision, computational efficiency, and interpretability.
• Neural networks (CNNs) are effective but often act as black boxes, limiting clinical utility.
• Interpreting complex patterns in brain scans remains a significant hurdle for widespread adoption.

Purpose of the Study:

• To introduce the adaptive spatial key-region identification (ASKRI) framework for enhanced brain image analysis.
• To improve the accuracy and interpretability of identifying regions of interest (ROIs) in brain scans.
• To reduce computational burden in deep learning models for neuroimaging without compromising performance.

Main Methods:

• ASKRI combines adaptive sampling (Shannon entropy), probability-mean-driven selection, and kriging for spatial uncertainty quantification.
• It integrates block-to-block kriging with statistical inference to interpolate CNN-derived classification performance.
• The framework is designed for seamless integration with convolutional neural networks (CNNs).

Main Results:

• ASKRI significantly reduces computational load for model training while maintaining predictive accuracy.
• The framework reliably identifies spatially consistent and biologically meaningful regions associated with aging in the TRACK-TBI dataset.
• ASKRI enhances both accuracy and interpretability in region of interest identification.

Conclusions:

• The ASKRI framework offers a novel, transparent, and resource-efficient approach to brain image analysis.
• It has the potential to advance diagnostic support in clinical settings by improving ROI identification.
• ASKRI addresses key limitations in current neuroimaging analysis, paving the way for more reliable clinical applications.