With the promise of deep learning in the analysis of brain imagery data significantly aiding in the detection, diagnosis, and potential solution to the treatment of brain disorders all relying on the power of deep learning, researchers, and clinicians with a great intent in enhancing the accuracy, efficiency, and objectivity of brain disorder detection [17] . Deep learning models excel at processing large volumes of brain scan image data, extracting meaningful features that a human eye may not ordinarily be able to identify.

Various architectures have been deployed to detect brain disorder, each with its own advantages and suitability for specific tasks. [18] deployed Convolutional Neural Networks (CNNs) in analyzing neurological imagery data in form of magnetic resonance imaging (MRI) and or positron emission tomography (PET) to uniquely detect brain disorders. Great success has been attained when capturing spatial patterns and has demonstrated impressive performance while detecting and segmenting brain tumors [19] . For brain disorders manifesting in 3D volumetric data, such as 3D MRI scans, 3D CNNs are used to learn spatial features and patterns, aiding in the detection of abnormalities [20] .

On the contrary, Recurrent Neural Networks (RNNs) have shown their effectiveness at modeling temporal dependencies, making them suitable for tasks that involve sequential data, such as time series analysis of brain signals [21] . With its ability in capturing temporal dependencies in brain signals, RNN’s have been applied for epilepsy detection and prediction of neurological disorders [22] . LSTMs are a type of RNN that can capture long-range dependencies in sequential data and have been used in brain disorder detection tasks like in the detection of Alzheimer’s disease (AD) using time-series datasets [23] .

Attention Mechanisms and Transformer models have shown relevance in capturing unique features from neuroimaging data and have been applied to tasks like detecting lesion and classification of diseases [24] [25] . They help in focusing on specific regions of brain images that are most relevant for detection, reducing the influence of irrelevant information and improving model accuracy [26] .

Autoencoders have been employed for feature extraction and dimensional reduction in brain imaging data, aiding in the identification of relevant features and anomalies [27] .

Using large scale datasets like such as ImageNet in pretraining deep learning models and then fine-tuning them for specific brain disorder detection tasks has become a common practice [28] . This approach, known as transfer learning, leverages the learned features from pretrained models, which can accelerate training and improve performance, especially when the labeled data is limited [29] [30] . By adapting pretrained models to brain imaging data, transfer learning facilitates the development of robust and accurate deep learning models for brain disorder detection.

Whereas transfer learning helps in scenarios of data paucity, other techniques like data augmentation have been adopted where data is artificially expanded on the training dataset to increase its diversity. Data Augmentation has found a vital role in brain disorder detection as it has been used in cases where data may be imbalanced and as such has limited labeled data. Common data augmentation techniques include rotations, translations, flips, and elastic deformations [31] [32] . These techniques help to enhance the generalization ability of deep learning models and improve their performance on brain imaging datasets.

Whereas deep learning models are known to have issues with Interpretability and explainability, it’s important to understand the processes that originate when deep learning models are making decisions if clinicians’ trust must be obtained to facilitate a smooth clinical adoption. Various methods, such as saliency maps, gradient-based attribution, and attention mechanisms, have been proposed to interpret and explain deep learning models predictive decisions.

Preprocessing is a core step that initiates data analysis and machine learning. It transforms raw data into information which can be used in developing predictive algorithms based on extracted features from these datasets. If a systematic process is followed to clean, transform, and select relevant feature domains from this data, hidden patterns in this data can be explored for an informative decision-making process. Various image preprocessing and feature extraction techniques and methods will be explored during this study. Several preprocessing techniques have been adopted and continue to play a critical role in accurately detecting other brain disorders. This study explores the common techniques used in preprocessing and feature extraction.

Skull stripping, also known as brain extraction, is a fundamental neuroimaging data preprocessing stage. It involves separating brain tissue from non-brain tissues for a given MRI scans. Non-brain tissue can be the skull, scalp, extracranial tissues, and other brain tissues with intent of isolating and focusing on the brain region of interest. Accurate skull stripping is crucial for subsequent analyses, such as brain segmentation, registration, and volumetric measurements.

Amongst the common skull stripping techniques used are Brain Extraction Tool (BET) which is a widely used and popular algorithm for skull stripping in neuroimaging. Being part of the FSL software suite, BET estimates brain tissue boundary and models out the actual brain tissue excluding other non-brain regions. [33] describes the BET as one that employs a principle of deformable surface modelling on an iterative deformed 3D surface mesh to properly fit the brain edges.

Other applications which perform the same role include FreeSurfer, a software used for structural MRI analysis. [34] describes FreeSurfer software as one that uses both intensity thresholding and surface-based modelling to separate non-brain tissue in each image generating a resultant high quality brain mask. Other notable applications like ROBEX (Robust Brain Extraction) [35] , ANTs herein as Advanced Normalization Tools [36] have found a wide applicability in skull stripping.

The act of image registration is a technique of wrapping some images to align their features in a perfect reference plane. This is usually done to ensure spatial correspondence amongst the images being preprocessed for group level analysis and comparisons to be done. Several methods have been used. Among the popular ones are FSL (FMRIB Software Library) FLIRT which according to [37] , FSL uses a transformed model to align a given image to a central reference plane for structural and functional neuroimages.

ANTs (herein as Advanced Normalization Tool) performs both skull stripping and image registrations tasks with a reach library with embedded algorithms like symmetric diffeomorphic registration. Because of its highly advanced image registration capabilities, ANT has found itself widely taken as an image normalization tool of choice in image registration [38] . Other notable applications like SPM (Statistical Parametric Mapping) which according to [39] is used for analyzing brain imaging, Elastix being a command line based software contains a wide range of transformation techniques like rigid, affine, and non-rigid enabling maximum customization of its features thus making it a dynamic tool for image registration [40] and finally ITK (Insight Segmentation and Registration Toolkit) an open source software package has found itself being widely adopted for medical image analysis by most of the research communities. This versatile nature according to [41] makes ITK a software package of choice in research.

The data collection process is prone to bring bias which can affect the overall quality of the data. The bias field correction technique is deployed to eliminate varying biases in the datasets specifically neuroimaging datasets, these inconsistences often corrected by shading off artifacts originate because of the imaging techniques used which can be biased by the pre-imaging clinical examinations miss directing the image acquisition process. Several methods have been deployed to correct these biases which include amongst others N3 (Non-parametric Non-uniform Normalization) Algorithm which according to use for bias correction in cases where the field is represented as a smooth spatial function. According to [42] , this bias correction technique has found a huge adoption rate in circumstances where images have features which may have inconsistencies that make them not clear for analysis. Other normalization techniques used in this area include algorithms like Polynomial Approximation with Linear Combinations of Exponentials (PALM) and non-parametric methods with specific algorithms like the GradWarp algorithm. Depending on the imaging data used and the study objectives, researchers can opt for a specific bias technique.

Intensity normalization is a preprocessing technique used to scale the image intensities of medical images, such as neuroimaging data from MRI (Magnetic Resonance Imaging) scans. The purpose of intensity normalization is to bring the image intensities into a standardized range, making the images comparable across different subjects or imaging sessions. This is particularly important in neuroimaging studies, where consistent intensity scaling is essential for accurate and reliable quantitative analyses.

To normalize images in such a manner that their mean and standard deviation give a zero, normalization technique of z-score has been widely adopted. [43] explains that as opposed to other normalization techniques like mini max and percentile normalization techniques, z-score transforms image pixels intensities to achieve a z-score with a mean and standard deviation of zero which method may become a method of choice depending on the characteristics an image possesses, and the requirements of a given research. Othe techniques like Motion Correction, Spatial Smoothing, Voxel-based Morphometry (VBM), Standard Space Normalization, me Series Preprocessing (fMRI) amongst others have been widely chosen for processing neuroimaging datasets. A summary of the key references has been outlined in Table 1 .

Several studies have deployed various feature extraction methods in classifying brain tissue.

PCA is a technique in dimensionality reduction that identifies key features (principal components) from a dataset and narrows down the data onto a lower-dimensional space while preserving as much variance as possible.

Studies [44] - [48] have deployed PCA feature extraction techniques and have achieved a significant performance.

LDA like PCA is a technique in dimensionality reduction, but it is often used in the context of classification problems. It targets projections with maximum separation amongst different class bounds.

[47] combined PCA and LDA to preprocess electroencephalography (EEG) data for brain activity classification. Neural Networks were added into PCA and LDA techniques for improved performance of the classification.

In computer vision, CNNs are often used to automatically learn hierarchical features from images. The activations of intermediate layers can serve as meaningful features for various tasks. The pivotal role CNN has shown in computer vision [48] with variants like MobileNet [49] and EfficientNet [50] and AlexNet [51] have led the adoption of CNN in image classification and objection detection tasks as well as in segmentation tasks.

CNNs are composed of multiple convolutional layers, pooling layers, and fully connected layers. The convolutional layers apply filters to the input data, scanning the data in both spatial and temporal dimensions.

Table 1 shows the references that discuss the various studies which have used various feature extraction techniques while Table 2 shows a detailed summary of the various feature extraction methods.

The following techniques have been widely adopted for feature extraction and a great success has been realized.

Wavelet transformation as applied in both signal and image processing, decomposes data into different frequency components, which frequency components will act as features for numerous analysis tasks.

Wavelet transformation has been applied in cancer detection [52] , image compression [53] , EEG Signal Analysis [54] , audio signal processing [55] and fault detection in machinery [56] and a great success has been achieved.

HOG is a feature extraction method for object detection and image recognition. It computes the distribution of gradient orientations in an image, which is used to capture the shape and texture of objects.

HOG has been deployed with a great performance attained in various aspects of object detection and recognition [57] pedestrian detection [58] , face detection [59] , traffic sign recognition [60] , human action recognition [61] , and hand gesture recognition [62] .

LBP is used in image analysis for texture classification. It encodes the relationship between the intensity values of a pixel and its neighbors, capturing texture information.

LBP has found applications in Face recognition [63] , texture classification [64] , object detection [65] , face analysis [66] , texture segmentation [67] and image retrieval [68] achieving an immense performance in specific tasks.

Model training is a process by which a portion of data is used to make the model learn from. The model learns patterns and relationships from a labeled dataset to make decisions (predictions) when exposed to new unseen datasets. The dataset is usually divided into two portions of either 80% training and 20% validation datasets or 75%: 20% respectively depending on the volume of available data. During the model building phase, the model is made to learn from the training dataset for purposes of learning while the validation set is used in the process known as hyperparameter tuning.

[71] introduced AlexNet, a DNN model having a network of five convolutional layers and three fully connected layers. AlexNet demonstrated great depth as compared to past models with a combination of over 60 million parameters. It makes use of ImageNet dataset which consists of 1,000 categories of 1.2 million high-res images, marking a substantial dataset improvement compared to past studies. Th key performance measure used is image classification accuracy which reduced the top-5 error rate to 15.3%. The AlexNet’s model success gave rise to a wide scale adoption of deep learning in computer vision, with a great possibility of extracting hierarchical features from raw images. The paper’s adoption of parallel processing, regularization techniques, and model visualization gave the researcher great strength, and its wide scale adoption came when AlexNet model was made Open source. During model training, the following steps are involved.

Clean and preprocess the dataset by applying the necessary transformations, such as text preprocessing or image normalization, to ensure consistent and well-formatted data. With data preparation being a foundational step in data science for purposes of ensuring quality and usability datasets, pandas’ official documentation by [71] offers a detailed guide in data preparation using python. This discussion is extended further by the authors distinguished works in python for data analysis. [73] in their publication on the art of data cleaning gives a detailed look at strategies to uncover challenges related to data quality.

Represent the data in a suitable format for the chosen algorithm. For text data, this could involve vectorization using techniques like TF-IDF or word embeddings. In image data, feature extraction might involve resizing, cropping, and color normalization. As much as feature engineering is an independent process in machine learning, it borrows a lot of aspects in relation to data preprocessing. [74] in the publication, “feature engineering for machine learning” dives to how features engineering gives strength into areas of data preparation.

Choose an appropriate algorithm or architecture based on the nature of the problem. For instance, CNNs which are often used for image-related tasks, RNNs used for sequence data tasks like textual datasets amongst other models depending on the nature of the available data. For simpler tasks, linear regression models can be adopted as discussed by [74] and for complex tasks like that used in ImageNet classification, deep Convolutional Neural Networks as discussed by [75] can be adopted.

Training dataset is fed into the chosen model and to minimize a chosen loss function, model parameters are iteratively being updated. The optimization process typically involves backpropagation and gradient descent with a notable stochastic Gradient Descent Algorithm [76] as used in the training of mostly Neural Network (NN) models amongst other models. During model training, it’s essential to select an appropriate loss function to determine the variation between the actual and predicted values. [77] discussed a popular loss function known as log-loss function (logistic regression). For regression tests, the Mean Square Error (MSE) loss function has been used to quantify the variation between the actual numerical value and the predicted value.

Hyperparameters such as learning rate, batch size, and network architecture are adjusted to optimize the model’s performance on the validation set.

Bayesian Optimization techniques have been widely adopted for hyperparameter tunning processes [78] . [79] introduces the grid search optimization technique where a range of values have been specified for each hyperparameter test and at each instance, the model is independently validated.

Model evaluation phase examines the model’s performance on new dataset to gauge how effective the model is in solving the desired goal and as well on its generalization ability. The evaluation process includes:

In model training, datasets are split into training and test datasets normally with either 75% to 25% or 80% to 20% proportions depending on the volume of the available data. A portion of the dataset (distinct from the training and validation datasets) is what is reserved as a test dataset. This data is used to simulate how the model will perform on real-world, unseen examples. [78] looks at data being split into the training dataset which seeks to make the model learn based on the features exposed to it, validation set examines model to make early assessments on model performance in one way fine tuning the hyperparameters.

Model Evaluation is a process of testing how well a trained model can accurately be able to perform its set task. Various metrics are used for model evaluation, and they depend on the nature of the problem being addressed. For classification tasks, metrics might include accuracy, precision, recall, F1-score, and ROC-AUC used specifically in cases where the dataset is imbalanced. [80] proposed using precision and recall while for regression tasks, metrics like Mean Absolute Error (MAE) and Root Mean Squared Error (RMSE) might be used with the intent of gauging the quality of the prediction. All these metrics are deployed with intent of identifying how well the model performs in real-world scenarios.

Model evaluation is used with a core aim of detecting overfitting where cross-validation techniques have been adopted as a means of reducing overfitting as discussed by [78] . It’s important to choose the right evaluation techniques depending on the data being used. For cases of MRI Brain Imagery for studies aimed towards the detection of Alzheimer’s Disease, recall can be used as opposed to accuracy given the false negatives and missed positives associated which are always associated with medical diagnosis.

The entire process of model training from the dataset creation to the final stage of trained model evaluation is shown on Figure 1 .

Visuals have been used to create a deeper understanding on the relationship between various features used in model training. Visualizations like confusion matrices, ROC curves, and precision-recall curves, provide a more comprehensive understanding of the model’s behavior.

In both evaluation and interpretation, iterative processes might be necessary. Models can be fine-tuned, hyperparameters adjusted, and preprocessing techniques optimized to achieve the best possible performance. Regular validation of models on new and diverse datasets helps ensure robustness and reliability in various scenarios. [81] discusses key visualization techniques used for model selection. [80] discusses advanced visualization techniques which are used for visualizing high dimensional data in lower dimensional spaces with methods such as Principal Component Analysis (PCA) and t-Distributed Stochastic Neighbor Embedding (t-SNE).