Convolutional neural networks (CNNs) are widely recognized for their high precision in image classification. Nevertheless, the lack of transparency in these black-box models raises concerns in sensitive domains such as healthcare, where understanding the knowledge acquired to derive outcomes can be challenging. To address this concern, several strategies within the field of explainable AI (XAI) have been developed to enhance model interpretability. This study introduces a novel XAI technique, GASHAP, which integrates a genetic algorithm (GA) with SHapley Additive exPlanations (SHAP) to improve the explainability of our 3D convolutional neural network (3D-CNN) model. The model is designed to classify magnetic resonance imaging (MRI) brain scans of individuals with Alzheimer’s disease and cognitively normal controls. Deep SHAP, a widely used XAI technique, facilitates the understanding of the influence exerted by various voxels on the final classification outcome (Lundberg SM, Lee SI. A unified approach to interpreting model predictions. In: Advances in Neural Information Processing Systems, 2017. 4765–74. https://doi.org/10.5555/3295222.3295230 ). However, voxel-level representation alone lacks interpretive clarity. Therefore, the objective of this study is to provide findings at the level of anatomically defined brain regions. Critical regions are identified by leveraging their SHAP values, followed by the application of a genetic algorithm to generate a definitive mask highlighting the most significant regions for Alzheimer’s disease diagnosis (Shahamat H, Saniee Abadeh M. Brain MRI analysis using a deep learning based evolutionary approach. Neural Netw. 2020;126:218–34. https://doi.org/10.1016/j.neunet.2020.03.017 PMID: 32259762). The research commenced by implementing a 3D-CNN for MRI image classification. Subsequently, the GASHAP technique was applied to enhance model transparency. The final result is a brain mask that delineates the pertinent regions crucial for Alzheimer’s disease diagnosis. Finally, a comparative analysis is conducted between our findings and those of previous studies.
Zahedipour et al. (Fri,) studied this question.