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

Brain image analysis presents significant challenges due to limitations in precision, computational efficiency, and interpretability. Although neural networks have proven effective for modeling complex patterns, they often function as black-box systems, making their predictions difficult to interpret and limiting their clinical utility. To address these challenges, we propose the adaptive spatial key-region identification (ASKRI) framework-a novel method to identify region of interest, which combines adaptive sampling based on Shannon entropy, probability-mean-driven selection, and spatial uncertainty quantified via kriging method. ASKRI integrates block-to-block kriging with statistical inference to interpolate CNN-derived classification performance, significantly reducing the computational burden of exhaustive model training without sacrificing predictive accuracy. Designed for seamless integration with convolutional neural networks (CNNs), ASKRI enhances both the accuracy and interpretability of ROI identification. Its effectiveness is demonstrated using the traumatic brain injury (TRACK-TBI) dataset, where ASKRI reliably identifies spatially consistent and biologically meaningful regions associated with aging. These results underscore the framework's potential to advance brain image analysis, while offering transparent and resource-efficient diagnostic support in clinical settings.