Model for autism disorder detection using deep learning

ABSTRACT


INTRODUCTION
Autism spectrum disorder (ASD) is a neurological aliment with different traits such as: difficulties with human engagement, cognitive impairments, and behaviors that are restricted and recurrent [1].Autism has been linked to both morphological and functional engagement abnormalities in numerous previous studies.The most often utilized imaging modality for evaluating anatomical anomalies is structural MRI (sMRI) [2], whereas functional MRI (fMRI) is by far the most widely used technique for investigating neural activity [3].
As per the World Health Organization, one out of every 160 children have ASD, and these children frequently suffer from sadness and anxiousness.Timely detection after birth is essential since it can benefit individuals with ASD in developing their interpersonal abilities and communication issues, as well as improve their overall quality of life.Early detection is important for controlling and treating this disease.Designing a model based on physiological or structural region connections in the brain is one of the most crucial jobs for detecting brain diseases such as epilepsy, Alzheimer's, and autism.
It is attributed to a slew of behavioral problems, some of which can become out of hand if the diagnosis is delayed [4].Despite the fact that symptoms appear in childhood, in most of the cases, a diagnosis is delayed [5].This is due to the fact that the current ASD diagnosing technique is entirely subjective and based on pre-defined questionnaires, which necessitates the pediatricians going through a children's cognitive background and conduct despite the fact that these procedures are extremely clear-cut, they are unquestionably  ISSN: 2252-8938 Int J Artif Intell, Vol. 13, No. 1, March 2024: 391-398 392 flawed, boundless, and may necessitate expertise that may not be accessible at the time at a large number of medical institutions.
With technological advancements, a great number of researchers are now analyzing an automatic computer assisted diagnosis of autism [6] as well as designing interactive technologies to assist in the treatment and rehabilitation of autistic people [7].Subjectivity would be reduced, and diagnostic repeatability and availability would be improved.It would also go a long way toward assuring early detection [8].Various neuropsychiatric and neurodegenerative conditions can be identified using magnetic resonance imaging (MRI).
In the past few years, machine learning (ML) techniques have shown significant promise in the analysis of pictures and videos.Particularly in a variety of fields, deep learning has achieved a lot of success.The success of deep learning has been aided by methodological developments as well as the widely accessible, massive datasets in the field of image processing [9].Unfortunately, neuroimaging does not operate in this way.Large datasets are hard to come by, if they exist at all.However, a number of organizations have worked to create a useable dataset by fusing neuroimaging datasets of persons with autism spectrum disorder (ASD) and people who are Typically developed (TD) from diverse locations across the world [10].About 2,150 participants, including ASD and TD, are included in the whole dataset, which comprises the first and second waves of data aggregation (ABIDE-I/ABIDE-II) [11].Although such a database exists, the enormous density of the input data makes it difficult to employ without some sort of preprocessing or pattern discovery [12].
In the proposed work, 3D neurological images having height, depth and volume have been considered to classify whether a person is autistic or not based on multiple brain scans in the fMRI format.An attempt has been made to study and evaluate the abnormalities in specific part of brain to narrow down our experimentation to final classification.Certain regions of brain have been targeted in order to frame region of interest and carry out the entire experimentation.A custom convolutional neural network (CNN) model is built to manipulate the classification of brain scans and determine whether or not the person is autistic.
The remaining portion of the paper has been presented using the following format: Section 2 presents a literature review of previous research articles on the topic of interest followed by problem formulation in section 3, and section 4 consist of detailed overview of the technology solution employed to achieve the intended goals.Section 5 delves into the technique employed, while section 6 outlines the review of the system's performance.The paper and recommended approach come to a close in section 7.

LITERATURE REVIEW
ABIDE, which is the result of a collaborative effort incorporating phenotypic data and neurobiological images acquired from 1,114 individuals, is the most widely used data-driven method for identifying biomarkers and diagnosing autism [13].Recently, there have been numerous attempts to diagnose ASD using fMRI and deep learning.Koyamada et al. [14] investigated a deep neural network (DNN) model to build a subject-transfer decoder.The authors build a decoder to examine various characteristics of every individual in the dataset utilizing primary sensitivity analysis (PSA).Their recommended neural system contains two hidden layers, a softmax output layer, and two visible layers that classifies brain activity from 499 people into seven human categories using the two hidden layers in the center.
Based on the examination of brain activation in response to a speech experiment, Haweel et al. [15] suggested a novel computer-aided grading system for infants and toddlers (between 1 and 4 years).Using functional connectivity characteristics of resting-state fMRI data, Subah et al. [16] proposed an ASD detection model.Two widely used brain atlases, Craddock 200 (CC200) and automated anatomical labeling (AAL), as well as two uncommon atlases, bootstrap analysis of stable clusters (BASC) and Power, are employed in their suggested model.
Masood et al. [17] used the deep convolutional neural network (CNN) inception model to train the model on spatial features, and recurrent neural network (RNN) to train the model on temporal features that are present in video sequences to prepare a sign language recognition model.Eslami et al. [18] presented a model based on the deep learning and machine learning technique using CNN and SVN classifier respectively for identifying attention-deficit/hyperactivity disorder (ADHD) and autism spectrum disorder (ASD) the model is used to identify the traits of autism using functional and structural state MRI of the subject.As a result, it was proved that deep learning (DL) model works in a better way than the proposed ML model.
Choi [19] used a variational autoencoder to transform multidimensional and high-dimensional data into two-dimensional features and showed a functional connectivity pattern related to Autistic children.Rad et al. [20] worked on stereotyped motor movements (SMM) in autistic people, which have an effect on learning and social skills including body swaying and intricate hand gestures.The multi-sensor accelerometer readings from SMM are processed by the CNN to extract various attributes.Havaei et al. [21] proposed a completely automated brain tumor segmentation approach based on CNN.
ISSN: 2252-8938  Devika et al. [22] suggested ASD detection model based on structural MRI For this purposes authors used only healthy patients' s MRI data and trained it using generative adversarial network (GAN) to recognize spatio-temporal patterns in skeletal brain connections.The model was contrasted with two alternative baselines, a more complex self-attention GAN and a more basic UNet.Furthermore, compared to cross-sectional data, longitudinal data and achieved 17-28% greater accuracy (one scan per subject).Using structural MRI data, Khadem-Reza et al. [23] suggested a novel method for differentiating ASD patients from controls.The accuracy of this method is increased by simultaneously utilizing the structural photos' volume and surface attributes.The authors' investigation included a number of ML and DL techniques, which achieved diagnosis accuracy of 86.29%, 71.15%, 86.53%, and 88.46%, respectively.Artificial neural network (ANN) offered the most accurate diagnostic possible.
As the preceding section demonstrates, there is a compelling need to examine the viability of utilizing DL-based models to detect ASD in the human population.The majority of the work thus far has relied on classic machine learning techniques, which have performance restrictions.The performance of a DL model was investigated for this study.

PROBLEM FORMULATION
ASD is a neuro-disorder that has long-term consequences for a person's interactions and communication with others.Because symptoms typically start to show between the ages of two and four, autism is usually detected and referred to as a "behavioral syndrome" [24].Experts believe that the ASD problem begins in childhood and continues through adolescence and adulthood.
The main goal of this study is to use convolutional neural networks of deep learning to categorize whether a patient is autistic or not.The brain scans have been classified and localized to the annotated areas to identify if the scan indicates autism or not.This study focuses on utilizing CNN with a neuroimaging dataset to automatically identify autism spectrum disorder (ASD) [25].The most popular resting-state functional magnetic resonance imaging (R-MRI) data from the autism brain imaging exchange (ABIDE), a multi-site dataset based on structural connectivity patterns, were used to identify ASD patients.The proposed method was successful in classifying both ASD and control subjects [26].

DATASET AND TECHNICAL APPROACH
ABIDE-I dataset is acquired from the sources to meet the research's purpose, and this dataset was then analyzed using numerous algorithms to determine its detection performance on 3D-image Data.Because the data contained a few non-contributing and subcategory attributes, it was preprocessed.Preprocessing is the process of transforming a data point before delivering it to the model.This is used to tidy up noisy or unreliable data so that it may be used for training and analysis.There is potential for enhancing the accuracy of existing approaches for identifying autism spectrum disorder using functional MRI to keep up with technological advances in the neuroimaging area.
The main goal of this research is to shorten the training period and improve the accuracy of mean time series signal-based ASD detection systems already in use.Table 1 presents the MRI data from 873 patients, including 405 individuals with ASD and 468 individuals without ASD, from the autism brain imaging data exchange (ABIDE) collection [12].ABIDE, the largest repository of data has amassed and disseminated 1,114 R-fMRI data, together with related structural MRI and observable attributes statistics from 541 people with autism spectrum disorders (ASDs) and 574 age-matched classic controls (6-65 years).
Based on assessments of 362 male ASD subjects and 405 male age-matched TCs, this resource is described and its applicability for furthering knowledge of ASD neurobiology is demonstrated.Whole-brain intrinsic cognitive control, as well as a variety of voxel-level inherent functioning brain architecture measurements, have been taken into consideration.Whole-brain analyses in the ASD literature harmonized seemingly contradictory themes of hypo-and hyperconnectivity; both were identified, but hyperconnectivity predominated, particularly in corticocortical and interhemispheric functional connectivity.

Data transformation
Data preparation is the process of transforming raw data into a useable and understandable format.Real-world data is frequently fragmented and unreliable because it contains so many inaccuracies and null values.Good pre-processed data is always the source of a favorable outcome.Several data pre-processing procedures are used to deal with missing and irregular data, such as processing null values, anomaly analysis, data denoising, subsampling (size and numerosity reduction), and so on.
The images are of the Neuroimaging Informatics Technology Initiative (NIFTI) form, which is commonly utilized in neurological scanning as shown in Figure 1.The volume in the MRI scans is also accounted for in the NIFTI files, and this volume is normalized for every scanning so that the scans can be processed correctly.After normalizing the volume, the scans are shrunk by normalizing the width and height.

PROPOSED METHODOLOGY
Resting-state functional magnetic resonance image (RS-FMRI) is utilized in a task-negative situation to map the brain and examine regional relationships.The vast majority of studies rely on having access to raw fMRI image data; however, these data are difficult to analyze and necessitate a long time, which makes them vulnerable to overfitting.A deep learning approach to diagnosing autism spectrum disorder using functional connectivity patterns derived from preprocessed FMRI data is described, taking into account the aforementioned problems.
A series of three-dimensional scans taken over time and evaluated by functional MRI scans are available and show a signal linked to brain activity (most often, the blood-oxygen-level-dependent signal, or BOLD signal).In this illustration, the individual is immobile inside the MRI scanner as a series of brain images showing the evolution of the BOLD signal strength are taken.A single preprocessed fMRI scan is therefore a four-dimensional time-series data collection with three spatial dimensions and time.Instead of analyzing the whole time series acquired from each brain voxel, certain brain regions of interest (as defined by the brain atlases) were evaluated.The photographs display the left brightly and temporal lobes, as well as the right frontal lobes.These specific regions have been chosen as the final test sites for our investigation.

Data preprocessing
Multiple augmentation approaches are applied to create synthetic data after preprocessing the scans and levelling their height, breadth, and volume.Below mentioned transforms are applied to the given dataset: 1. Flipd → For flipping when working with a dictionary.2. Rotated → To apply rotations to a dictionary.RandGaussianNoised → To apply a gaussian noise to the dictionary. 5. Finally, there is RandAffined → Actually, this function can perform multiple transformations at the same time; it is used to perform translations, but it can also be used to perform rotation if the Rotated function is not supposed to be used.

Training and testing model
The complete dataset was divided into two halves with a 75:25 ratio, one for training and the other for testing.For cross-validation purposes, the training set has been separated into two halves.The training set is split in half, with one half used for training and the other for validation, with a 75:25 split.The final train, test, and validation sets that are utilized for classification are then separated and used for further processing.

Model training
The next step in the pipeline is to proceed to neural network building after augmenting the data.Threedimensional CNN layers will be used in the neural network.The operations of 3D CNN layers are the same as 2D CNN layers, only the filter moves in three directions instead of two.Parametric rectified linear unit (PRelu) is used as the activation function, and 3D pooling layers are used in the network.The image is processed and passed through multiple layers before being classified into one of two classes as shown in Figure 2. The model uses pooling and convolutional neural network layers.Now that the features have been provided, the CNN layer filters the image in a 3×3 manner.Reducing the image's pixel size is the first step.This enables the image to be processed more quickly.The model's performance is improved by employing data augmentation techniques like picture rotation, image scaling, cropping, padding, translation, and so on to create fresh and varied cases for the model to train on.A sequential model is built utilizing 3D convolution layers, max pool layers, and the activation function "PRelu" after data augmentation is supplied to the training set.A PReLU, or parametric rectified linear unit, is a generalization of the classic restored unit with a gradient for negative values as shown in ( 3) and (4).Formally, Now that the image has been loaded, the first Conv3D layer analyzes it using a 3×3 filter to search for distinctive features.After extracting both sharp and smooth features, the pooling layer modifies the weights based on the feature map's greatest value.The activation function PRelu only accepts positive values; all other inputs are 0. As a result, the model's effectiveness is increased because only chosen features are transferred to the following layer.PRelu enables the model to train more quickly and prevents the fading gradient issue.Before the model is saved and exported for use, Dense layers are used to correct the receiving vector's dimension as needed after all of the layers have been applied and the resultant picture vector has been flattened into an array.The model summary is shown in

RESULTS AND EVALUATION
In today's day and age, ASD is a fast-expanding disorder that affects people of all ages.Early detection of this neurological issue can aid in the subject's overall health and well-being.Early diagnosis based on a variety of medical and physiological factors now looks to be feasible, thanks to the rising use of deep learningbased models to forecast various diseases in humans.
CNN is frequently used today to categorize datasets.In this study, we developed a CNN model for automated ASD identification using the ABIDE I dataset.From the ABIDE I dataset, we used preprocessed neuroimaging data for our investigation.The 1,114 patients in the ABIDE I dataset-of which 541 had ASD and 574 were typical controls-were preprocessed down to 875.
A convolutional neural network-based model has been integrated into this work in order to diagnose persons with autism based on their facial expressions.Multiple convolutional layers are used to create a specific DL model.The collected findings support the model's sound performance.As a result, the method described in this study has, to the best of our knowledge, produced the highest accuracy when using the ABIDE I dataset.

Evaluation metric
Certain assessment criteria have been incorporated in order to evaluate the performance of the proposed model in classifying the patients as autistic or non-autistic.For classification purpose, the model is evaluated based on accuracy and F1-score.Accuracy (3) determines the ratio of successful assertions made to the total number of instances provided as input.The F1-score, which is defined as the harmonic mean of the recall and precision of the model, is a way of combining the model's performance is depicted in (4).The result is shown in Table 3.The results have a high level of significance as a result of everything.The results of the trials indicate that the proposed model is capable of accurately predicting and classifying ASD.Taking into account all of the data and model findings, the proposed technique is determined to be far more efficient than alternative frameworks now in use.As a result, the proposed methodology can be described as a suitable model for distinguishing between autistic and non-autistic people.

CONCLUSION
In this study, ASD was predicted using a deep learning technique based on multisite resting-state fMRI.ASD identification is a challenging task due to the lack of a standard modeling solution and the wide variety of current practice.The results show that our model outperformed the previous best accuracy on this dataset, with an average accuracy of 92.45 percent using the test data.For novel models, it has been found that a CNN model with lower dimensionality is more effective and has lower operational expenses.Because of this, it is possible that our model will be able to train with fewer features and still be as accurate as the best models.Predicated to the results obtained by the experimentation on this problem, the CNN classifier showed significantly reliable performance when all of the feature's characteristics were included after missing values were handled.In comparison to previous investigations, the current study showed a considerable improvement in performance.Despite this, there are several restrictions that must be addressed.

Int 3 .
Model for autism disorder detection using deep learning (Anshu Sharma)395 Zoomd → To apply a zoom.4.

Figure 2 .
Figure 2. Proposed deep neural network architecture to predict ASD

Table 1 .
Phenotypic information summary of the participants from the ABIDE dataset Figure 1.3D images in ABIDE dataset

Table 2 .
Convolutional neural network model

Table 3 .
Evaluation result of the CNN based classification modelAnother criterion taken for the is the mean squared error rate and mean absolute error rate, which received a score of 0.30 and 0.29 respectively.For all of the assessment indicators taken into account, the technique's results are substantial as shown in Table4.The findings of the preceding investigation led to the following conclusions: The prediction model's accuracy is 93.41 percent after applying custom CNN architecture to the 3D brain dataset acquired from the source, while F1-Score is 0.93.−Thescoreachieved for the Mean Squared Error is 0.29.−Thescore achieved for the Mean Squared Error is 0.30.
Model for autism disorder detection using deep learning (Anshu Sharma) 397 −

Table 4 .
Evaluation result of the localisation of the brain data