2.1. Alzheimer’s classification

Numerous studies have presented diverse neural network architectures for the purpose of classifying medical images, as this task holds significant benefits for the development of XAI techniques. Consequently, numerous studies have employed image classification techniques for the purpose of diagnosing Alzheimer’s disease [20–22]. In recent years, transfer learning has gained prominence due to its capacity to optimize time and resources by obviating the need to train several machine learning models from the ground up in order to do similar tasks. One example involves the proposal of a technique that employs deep neural convolutional networks, transfer learning, MRI inside the Visual Geometry Group (VGG) 16 and VGG 19 models [23]. This study additionally demonstrated that Alzheimer’s disease (AD) may be classified into four distinct groups, which are commonly recognized by medical professionals, through the use of computer algorithms. The authors of the study [24] present a deep feature-based methodology for the identification of Alzheimer’s disease stages. In order to accomplish this task, the initial layers of a pre-existing AlexNet model are utilized, and afterwards, the deep features are learned using a convolutional neural network. In this study, the deep features that have been extracted are subjected to classification using several machine learning approaches, such as K-nearest neighbor (KNN), random forest (RF), and support vector machine (SVM). Ebrahimi-Ghahnavieh et al. [25] present a strategy that combines CNN and recurrent neural networks (RNN) with the aim of improving classification performance. The initial step involves the partitioning of every three-dimensional MRI image into separate two-dimensional images. Subsequently, each of these two-dimensional images undergoes a distinct retraining process utilizing a distinct two-dimensional CNN. In order to build a significant correlation between these sequences and facilitate the process of decision-making, the Long Short-Term Memory (LSTM) model is employed to aggregate these inputs. The limited availability of comprehensive medical imaging training data is a challenge to the performance of CNNs and has the potential to result in overfitting. The study conducted by Acharya et al. [26] intended to employ transfer learning models, including VGG16, ResNet-50, and AlexNet, along with convolutional neural networks, to classify MRI scans of patients with Alzheimer’s disease into various classifications. Furthermore, this approach successfully classified several stages of Alzheimer’s disease, including mild cognitive impairment, mild Alzheimer’s disease, moderate Alzheimer’s disease, and severe impairment, with a high accuracy rate of 95.70%. Recent studies have further expanded the scope of Alzheimer’s Disease (AD) classification by incorporating explainability and diverse data types. For instance, a study utilizing a large clinical dataset from the National Alzheimer’s Coordinating Center employed traditional machine learning models such as SVM and RF with explainability techniques like SIRUS, CAR, SHAP, and LIME, achieving high accuracies (up to 99% for binary classification) [27]. Another study proposed a multilayer multimodal detection and prediction model using MRI, PET, and genetic data with RF classifiers, achieving accuracies of 94.4% for AD detection and 86.8% for progression prediction, enhanced by SHAP and Decision Trees for explainability [28]. Additionally, a deep transfer learning approach using ensemble models (e.g., VGG16, DenseNet) on MRI data achieved 96% accuracy, employing XAI techniques like saliency maps and GRAD-CAM, providing a direct comparison to GASHAP’s performance [29]. Recent advancements include multimodal deep learning with MRI, PET, and genetic data, achieving high diagnostic accuracy [30,31], and graph neural networks for connectivity-based AD classification [32]. Transfer learning with VGG19 and DenseNet has been explored, with SHAP and LIME enhancing interpretability [33], while random forest models predict (Mild Cognitive Impairment (MCI) to Alzheimer’s Disease (AD) progression (MCI-to-AD)) transition using multimodal data [34]. Three-class classification (control, MCI, AD) using deep learning and SHAP further refines diagnostic granularity [35].

Convolutional Neural Networks (CNNs) have demonstrated remarkable capability in automatically learning spatial and hierarchical features directly from medical images, eliminating the need for handcrafted feature extraction that traditional machine learning algorithms such as Support Vector Machines (SVM), Random Forests (RF), and k-Nearest Neighbors (k-NN) typically require. This ability allows CNNs to capture subtle structural and textural patterns within MRI data that may not be detectable through conventional techniques. Furthermore, CNNs provide scalability to high-dimensional input data, enabling end-to-end optimization for disease classification tasks such as Alzheimer’s diagnosis.

Despite these advantages, CNN-based approaches also present certain limitations. They often demand substantial computational resources and large annotated datasets for effective training, which can be restrictive in medical imaging domains where data acquisition is expensive and labeling is time-consuming. Additionally, CNNs have been criticized for their limited interpretability, as the internal decision-making process is not easily understandable to clinicians. This “black-box” nature underscores the necessity for XAI frameworks such as GASHAP, which aim to enhance model transparency while maintaining diagnostic accuracy.

A comparative summary highlighting the relative strengths and weaknesses of CNNs versus conventional machine learning methods is provided in Table 8, demonstrating how deep architectures improve representational power at the cost of increased complexity and reduced interpretability

2.2. XAI methods

A recent systematic review of XAI in Alzheimer’s disease classification analyzed 37 studies, covering frameworks such as LIME, SHAP, GradCAM, and Layer-wise Relevance Propagation (LRP) [36]. It highlighted critical gaps, including the need for validation with medical professionals and comprehensive datasets, which the GASHAP method addresses through its focus on MRI and precise brain region identification. In recent times, a considerable amount of scholarly attention has been directed toward the development of strategies aimed at enhancing the transparency of neural network models. This has been achieved through the integration of many algorithms or the independent use of diverse methodologies. LRP, initially proposed by Bach et al. [37], is an algorithm that is widely employed and considered to be highly intriguing. The utilization of LRP for the visualization of the CNN model, as described by Böhle et al. [38], is employed in the context of Alzheimer’s disease (AD) classification using MRI. In order to highlight the benefits of the LRP algorithm, a comparison is being made between this algorithm and the guided backpropagation (GB) technique, which is a gradient-based method. Ultimately, the researchers revealed significant cerebral regions that are crucial for the diagnosis of Alzheimer’s disease. The output of a CNN model can be visualized as a heatmap by employing Shapley additive explanations (SHAP), a technique that bears resemblance to LRP (Layer-wise Relevance Propagation) as proposed by Shapley [39]. In the study conducted by van der Velden et al. [40], a 3-dimensional regression CNN was utilized for the purpose of evaluating volumetric breast density. The primary benefit of the methodology suggested in this research article lies in its ability to circumvent the expensive process of conducting 3-dimensional voxel-level segmentations of the breast and fibroglandular tissue. One notable feature of the methodology outlined in this study is the use of a density score. In cases where radiologists encounter uncertainty regarding the density score, they may turn to the Shapley additive explanations for guidance. Seo et al. [41] introduced a regional, multi-scale, model-agnostic approach aimed at enhancing the interpretability of deep neural networks through the use of visualization techniques. This approach utilizes a region-based methodology to estimate the individual contribution of each feature towards the ultimate outcome. This estimation is achieved by omitting a specific region, which is generated through the sampling of a normal distribution formed based on the boundary prior. This methodology involves the segmentation of images into distinct layers, thereby facilitating a more precise and comprehensive elucidation of the overall shape. The degree of detail in the explanation is enhanced through the process of segmenting the photos into higher tiers. The study conducted by Sudar et al. [42] focuses on the classification of MRI Alzheimer’s brain images using a CNN architecture called VGG-16. The experiment uses the LRP technique to interpret the features utilized by the model for prediction. The collection has around 6400 photographs that have been categorized into four separate classifications. The results of this study suggest that the application of the VGG-16 model and the LRP approach is effective in the classification of images associated with Alzheimer’s disease. Furthermore, these findings provide significant insights into the unique attributes that are employed for precise forecasting. The genetic algorithm is a metaheuristic algorithm that operates on a population and can be employed for the interpretation of deep neural networks. In Shahamat and Abadeh [2], we focused on the identification of significant brain areas associated with autism and Alzheimer’s disease through the use of a genetic algorithm. Initially, a CNN is trained. Subsequently, genetic algorithm properties, including population initialization, fitness evaluation, crossover, and mutation, were established. The chromosome encoding utilized in this approach is derived from the Harvard brain atlas, where each gene within a chromosome is assigned a random value from the set S = {0, 1, 2, 3}. The brain regions that have been acquired are identified as crucial brain regions in the diagnosis of Alzheimer’s disease and autism. In Shojaei et al. [43], which is another important work in the field of XAI, we presented a research inquiry that explores the utilization of deep neural networks (DNNs) and CNNs for the purpose of diagnosing Alzheimer’s disease (AD) by analyzing MRI data. The integration of a genetic algorithm-based occlusion map technique with backpropagation-based explainability approaches was employed to identify the specific brain regions that significantly contribute to the output of the network. The extracted regions underwent a comparative examination with prior studies, which revealed their effectiveness in diagnosing Alzheimer’s disease. This conclusion was supported by professionals in the field of Alzheimer’s research. The model achieved an accuracy of 87% during the 5-fold cross-validation procedure, indicating a performance level comparable to that observed in prior research in a similar context. In addition, we achieved a validation accuracy of 93% using only 29 brain regions and implementing the lrp_z_plus_fast explainability approach. This study demonstrates the efficacy of a merged approach for the diagnosis of Alzheimer’s disease (AD) and contributes to the existing body of literature in this field. Ahn et al. [44] developed a genetic algorithm with the purpose of extracting meaningful features from the complete dataset. The researchers in this study provided a traffic classifier based on ResNet and afterwards discussed the importance of each feature by establishing the dominance rate. The proposed methodology employs a genetic algorithm to generate a feature selection mask. The classifier assesses the performance of each mask and selects those with the most favorable fitting scores. Shad et al. [45] focused on using the LIME framework for XAI in Alzheimer’s disease (AD) diagnosis. They aimed to improve the interpretability of deep learning models by providing explanations and visualizations using the LIME framework. Their hybrid dataset, which included MRI imaging data, allowed them to classify patients into different stages of dementia and distinguish AD patients from normal individuals. The researchers highlighted the importance of transparency in the medical field and demonstrated the potential of the LIME-based XAI framework in enhancing AD diagnosis by aiding in early detection and treatment planning. Bordin et al. [46] addressed XAI in Alzheimer’s disease (AD) classification using deep learning models. They trained a DL model on MRI scans of individuals with AD and healthy controls, achieving good performance. The occlusion sensitivity method was used to understand the DL model’s decision-making process, revealing the importance of white matter hyperintensity (WMH) lesions in AD identification. DL models can effectively utilize clinical information, particularly WMHs, as a biomarker for AD dementia. XAI methods provide insights and enhance trust in DL approaches in clinical settings. Recent advancements in XAI for AD classification emphasize SHAP for interpreting deep learning models, with studies using SHAP to identify key MRI-based biomarkers [35,47–51] and multimodal features like cognitive scores and genetic data [30,31,34,52,53]. Systematic reviews highlight the role of SHAP and LIME in AD detection, addressing clinical validation needs [54]. SHAP has been applied to graph-based models for brain connectivity analysis [55] and multitask deep learning for MCI-to-AD progression [56]. Alternative XAI methods, such as Grad-CAM and LIME, provide visual and local explanations for 3D-CNNs and transfer learning models [33,57,58], while LRP identifies critical brain regions [59]. Integrated SHAP-LIME frameworks offer hybrid interpretability [60], enhancing diagnostic transparency. The GASHAP method, combining genetic algorithms with SHAP, aligns with these trends, particularly in SHAP-driven region-based explainability [51].

3.1. Ethical considerations

This study utilized de-identified MRI data from the Alzheimer’s Disease Neuroimaging Initiative (ADNI), with approval from the ADNI Data and Publications Committee and relevant institutional review boards [19].

3.2. Data preprocessing

The MRI data used in this study were obtained from the ADNI repository (https://adni.loni.usc.edu/). Following the protocol proposed by Shahamat and Saniee Abadeh [2], a total of 140 T1-weighted MRI scans were selected, consisting of 70 scans from patients diagnosed with Alzheimer’s disease (AD) and 70 scans from cognitively normal (CN) controls. The demographic variables (age and gender) were approximately balanced across the two groups to minimize sampling bias. All MRI volumes were preprocessed using the FSL toolbox. Each image was reoriented to the standard MNI152_T1_2 mm space, skull-stripped, and intensity-normalized. Subsequently, the images were cropped to an 80 × 80 × 80 voxel subvolume centered on the brain to remove redundant background regions and reduce computational cost. The dataset was divided into training, validation, and testing subsets with a 70–15–15% ratio, maintaining class balance across subsets. In addition, a 5-fold cross-validation procedure was implemented to ensure the robustness and generalization of the model. Each MRI volume was normalized to zero mean and unit variance before being fed into the CNN model. Implementation details: All experiments were implemented in Python using the TensorFlow deep learning framework within the Google Colab environment. Training was performed on a GPU provided by Colab. Random seeds were fixed to ensure reproducibility. All analyses and training sessions were conducted under a consistent software setup.

3.3. CNN classification

The CNN architecture used in this study followed the design proposed by Shahamat and Saniee Abadeh [2]. The model consisted of three 3D convolutional blocks, each followed by a rectified linear unit (ReLU) activation and a 2 × 2 × 2 max-pooling layer. The first, second, and third convolutional layers employed kernel sizes of 5 × 5 × 5, 3 × 3 × 3, and 3 × 3 × 3, respectively. The output of the final convolutional block was flattened and connected to two fully connected layers containing 1024 and 2 neurons, respectively. A softmax activation function was applied in the output layer to produce class probabilities for Alzheimer’s disease (AD) and cognitively normal (CN) subjects.

The network was trained from scratch using the Adadelta optimizer with an initial learning rate of 0.05 and a decay rate of 0.95. The categorical cross-entropy loss function was used to measure prediction error. The model was trained with a batch size of 32 for up to 30,000 iterations (approximately 100 epochs). An early stopping criterion was applied when the validation loss did not improve for 10 consecutive epochs.

The model achieving the highest validation accuracy was saved for further analysis. This trained CNN was subsequently used to compute voxel-level SHAP values, which served as input to the genetic algorithm phase for generating interpretable regional masks (Table 1).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 1. The 3D-CNN architecture.

https://doi.org/10.1371/journal.pone.0337800.t001

3.4. Calculating shapley values

The SHAP methodology is a game-theoretic strategy that provides a concise approach for interpreting machine learning models, especially those with a black-box nature [1]. The SHAP framework offers implementations that can be applied a variety of machine learning models, such as Kernel SHAP, Tree SHAP, and Deep SHAP. The SHAP approach produces a unique and well-defined solution that exhibits three desirable characteristics: local accuracy, missingness, and consistency.

Property 1 (local accuracy)

When the explanation model agrees with the original model, The local accuracy equation is shown in Equation 1.

Equation 1. Local Accuracy:

(1)

Property 2 (Missingness)

Features that are constrained by missingness have no attributable influence. This desirable quality is displayed by Equation 2.

Equation 2. Missingness:

(2)

Property 3 (Consistency)

Equation 3 explains the consistency property. Let and denote setting . For any two models and if:

Equation 3. Consistency:

(3)

for all inputs , then .

Deep SHAP is a viable approach for elucidating the workings of deep neural network models. In this study, we employed a novel approach to increase transparency in black-box models by identifying crucial voxels in brain MRI images. This step of the GASHAP process results in the generation of heatmaps corresponding to each MRI image, as depicted in Fig 2. In the subsequent section, we describe the use of Shapley values in the generation of chromosomes for the genetic algorithm.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 2. 10 samples of MRI images and their corresponding heatmaps.

https://doi.org/10.1371/journal.pone.0337800.g002

3.5. Genetic algorithm

In the previous stage, the objective was to identify the voxels within the MRI images that exerted the most significant influence on the final outcome. To facilitate interpretation, the Harvard-Oxford cortical and subcortical structural atlas was used to delineate a total of 96 distinct brain regions, as shown in Table 2. This approach enabled analysis of results at the brain region level. Masks were generated using the previously computed Shapley values, as depicted in Fig 7. However, it is important to note that the SHAP technique, which provides explanations at a local level, may not yield precise outcomes when averaging Shapley values across all MRI scans. To address this limitation, a technique was utilized to convert each heatmap into a chromosome. The implemented genetic algorithm consists of five distinct steps, outlined below:

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 2. Predefined names of brain regions and their corresponding IDs.

https://doi.org/10.1371/journal.pone.0337800.t002

1. 1. Initial population
2. 2. Fitness function
3. 3. Selection method
4. 4. Crossover
5. 5. Mutation

This section will outline the stages involved in the implementation of the genetic algorithm. The fitness function, crossover, selection strategy, and mutation employed in this study closely adhere to the approach described in our previous study [2]. However, we have made modifications to the initialization of the population based on the heatmaps generated by Deep SHAP.

3.5.1. Initial population.

3.5.1.1. Chromosome Encoding: Based on the findings presented in the Harvard-Oxford Brain Region Atlas (Table 2), it was observed that there were a total of 96 distinct brain regions. Consequently, in our study, we designated each chromosome as a vector with a length of 96. The value assigned to each element of the vector is determined based on the corresponding values derived from the significance of every value within set S is provided in Table 3.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 3. Brain region importance.

https://doi.org/10.1371/journal.pone.0337800.t003

3.4.1.2. Score calculation: For each heatmap, the brain regions described in the Harvard-Oxford Atlas must be used to assign values from Table 3 to the genes on each chromosome. Subsequently, a score is calculated for each region by employing the mathematical formula presented in Equation 4. The diagram in Fig 3 illustrates the overall process of score computation. It should be noted that represents the highest value among the SHAP values inside the designated region.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 3. The process of score calculation for a heatmap.

https://doi.org/10.1371/journal.pone.0337800.g003

Equation 4. Score Calculation Formula Based on SHAP Values:

(4)

In the course of the investigation, it becomes clear that a location with a higher score is indicative of a higher level of criticality. In order to preserve the accuracy of the model, the equation eliminated zeros and negative SHAP values. The numerical value assigned to each region, as indicated by Equation 4, spans the range of zero to one. Following this, a numerical value from the set S = {0, 1, 2, 3} was allocated to each brain region, taking into account the outcomes of the preceding computation. The resulting brain masks are illustrated in Fig 4. The scores were arranged in ascending order, and a value of 0 was assigned to the lowest 30% of the group based on their scores. In a similar manner, numerical values of 1 were allocated to scores falling within the range of 30–50%, values of 2 were assigned to scores falling within the range of 50–70%, and values of 3 were assigned to scores falling within the range of 70–100%. The brain region exhibiting a gene value of 0 was omitted during the generation of the brain mask. In the event that a gene possessed a value of 1, morphological erosion was employed to diminish the dimensions of the associated brain region, given its lower score and potential scarcity of crucial voxels. Conversely, a gene exhibiting a value of 3, which is assigned a higher score according to SHAP, may possess a greater number of valuable voxels. Consequently, the associated region was expanded through the use of morphological dilation. Fig 5 illustrates the developed chromosome, while Fig 6 depicts the transformation outcome of the generated chromosome into a brain mask. The schematic representation of the chromosomal generation process in GASHAP is depicted in Fig 7.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 4. The Process of Applying Ranges to the Score Array for Creating Chromosomes.

1. After calculating the SHAP values and applying Equation 4, a score is assigned to each region in every MRI image. Consequently, each heatmap contains an array in which the scores of each region are stored. 2. Each area is placed in a range according to its score, and a value from the score range table is assigned to it. 3. That is, the first 30 percent that have the lowest score are assigned a value of 0, those that are between 30 and 50 percent are assigned a value of 1, the areas that are between 50 and 70 are assigned a value of 2, and the rest are assigned a value of 3.

https://doi.org/10.1371/journal.pone.0337800.g004

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 5. An example of a generated chromosome with 96 genes.

https://doi.org/10.1371/journal.pone.0337800.g005

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 6. The corresponding regions of the chromosome produced in Fig 5.

https://doi.org/10.1371/journal.pone.0337800.g006

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 7. Overall diagram of chromosome-generating flow.

https://doi.org/10.1371/journal.pone.0337800.g007

3.5.2. Fitness function.

As stated previously, each chromosome represents a distinct manifestation of the brain mask. The fitness function is defined according to Equation 5 to evaluate the efficacy of the generated chromosomes. The evaluation of the precision of each scan involves an initial conversion of each chromosome into a brain mask, which is subsequently used to process all input MRI pictures. The fitness value is subsequently computed by taking the reciprocal of the quantity representing the count of pertinent brain regions and combining it with a weighted mean of the model’s precision. The equation presented in question 5 provides a comprehensive definition of the fitness function: it represents the accuracy of processed input scans, the inverse of the non-zero elements, and variables that optimize both accuracy and the number of selected brain regions.

Equation 5. Fitness function:

(5)

Equation 6. Mask accuracy where TP, FP, TN, and FN denote true positives, false positives, true negatives, and false negatives:

(6)

Equation 7. Inverse number of selected brain regions:

(7)

The GA was employed to optimize the SHAP-based regional masks following the procedure proposed in our previous work [2]. Each chromosome encoded a binary vector representing candidate brain regions, and the fitness function was computed as a weighted combination of classification accuracy and mask compactness.

The GA was configured with a population size of 200, a maximum of 500 generations, a crossover probability of 0.4, and a mutation probability of 0.6. Tournament selection was used to choose parents for reproduction. The evolution was terminated either when the maximum number of generations was reached or when the improvement in mean fitness between consecutive generations fell below 0.001.

The weighting coefficients α = 0.025 and β = 0.975 in the fitness function were adopted from Shahamat and Saniee Abadeh [2], where they were shown to provide a good trade-off between accuracy and interpretability. The small α value gives a minor penalty to mask size to prevent excessive pruning of relevant voxels, while the large β value emphasizes classification accuracy to maintain diagnostic reliability. Preliminary experiments in this study confirmed that these parameter values produced stable convergence behavior and high validation accuracy.

3.5.3. Selection method.

The parental candidates for the subsequent generation are chosen from the population as the most appropriate answers. Various selection strategies can be employed to accomplish this task. In the present study, the roulette wheel method has been employed. Equation 8 can be used to compute the likelihood associated with the selection of each individual member.

Equation 8. The probability of selecting ith member of population:

(8)

4.1. Discussion on brain regions discovered for Alzheimer’s disease

Table 6 presents the brain regions identified from the first and second experiments. The genetic algorithm was initially executed using the specified properties outlined in the first experiment. The resulting brain areas were then assigned to the set . Following that, the genetic algorithm was re-executed using the specified properties outlined in the second experiment. The resulting brain regions were then assigned to the set . is the set that can be defined as the set that contains elements that are common to both sets.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 6. Final brain regions identified by the GASHAP framework.

https://doi.org/10.1371/journal.pone.0337800.t006

The individuals comprising this are shown in Fig 16. The primary objective of the intersection of sets was to improve accuracy and reveal more exact findings. To achieve a thorough understanding of a heatmap, it is sometimes necessary to evaluate it at different scales. For example, through careful examination of the image at a reduced scale, one can perceive finer details, while analyzing the image at an enlarged scale offers a more comprehensive understanding of the sent information. To conduct a thorough analysis of our hypothesis, we employ the use of two experiments and their intersection. This approach allows for a comprehensive examination of the hypothesis across both experiments, yielding a more robust outcome.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 16. .

https://doi.org/10.1371/journal.pone.0337800.g016

The results of the GASHAP study have been previously published in several papers. Of particular relevance, a substantial portion of the identified brain regions demonstrated a gene value of 3, highlighting the robustness of this methodology. The study demonstrates high accuracy, as evidenced by its performance on training and test datasets, its convergence rate, and the number of regions identified in the initial stages. A comprehensive list of the identified brain regions is presented in Table 6, along with references to previous studies that have confirmed their validity. The brain regions identified in the study include:

• Left cingulate gyrus, posterior division
• Left frontal pole
• Left Inferior Frontal Gyrus, pars opercularis
• Left Lateral Occipital Cortex, superior division
• Left Parahippocampal Gyrus, anterior division
• Left Superior Temporal Gyrus, posterior division
• Left Supramarginal Gyrus, anterior division
• Right Parahippocampal Gyrus, anterior division
• Right Superior Frontal Gyrus
• Right Superior Temporal Gyrus, posterior division
• Right Supramarginal Gyrus, anterior division

The final mask visualization generated by the study is shown in Fig 17. The corresponding chromosomes of the identified brain regions are displayed in Fig 18, with the coloring scheme explained in Fig 19. The details of the final results, along with relevant previous research, are presented in Table 6. The brain regions identified by the proposed GASHAP framework are summarized in Table 7, which lists the regions most relevant for Alzheimer’s disease classification (Fig 20). A comparison of the GASHAP framework with previous GA-based explainability studies is presented in Table 8.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 7. Extracted brain regions from GASHAP.

https://doi.org/10.1371/journal.pone.0337800.t007

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 8. Comparing GASHAP with previous studies.

https://doi.org/10.1371/journal.pone.0337800.t008

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 17. Discovered brain regions and final generated mask by GASHAP.

https://doi.org/10.1371/journal.pone.0337800.g017

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 18. Final discovered chromosome.

https://doi.org/10.1371/journal.pone.0337800.g018

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 19. A guide for Fig 18 coloring.

https://doi.org/10.1371/journal.pone.0337800.g019

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 20. Harvard-Oxford cortical and subcortical structural atlas.

https://doi.org/10.1371/journal.pone.0337800.g020