Automated Parkinson's Disease Detection: A Review of Techniques, Datasets, Modalities, and Open Challenges

Substantial research work is being done in the field of PsD detection and management. Although these publications cover a significant amount of work on its symptoms, modalities, detection, cure, management, etc., but some crucial aspects have been overlooked. Table 2 presents the relative contrast of the proposed survey with the state-of-the-art surveys on PsD detection. No existing survey has covered the scope of EI in the detection of the disease, and also only a few modalities are discussed. The motivation of the present paper is to present a dedicated review on the latest PsD techniques, most-used datasets, modalities, and open challenges. It offers a comprehensive survey of almost all the techniques of AI in PsD detection, from both application and methodology views.

In 2020, M. Noor et.al. and H. Loh et al., in 2021, highlighted that the convolutional neural network (CNN) model is widely used and achieves better accuracy in PsD detection. H. Khachnaoui et al., A. ul Haq et al., and M. Alzubaidi et al. presented a correlative analysis of ML as well as DL models and concluded that DL performed better than ML. M. Alzubaidi et al. also identified the best-performing DL models as well as the datasets. According to him, DNN achieved phenomenal results on the VGRF dataset, CNN and FNN achieved outstanding result for image classification, the neural networks gave promising results on voice data, and LSTM outperformed on EEG data. Tanveer et al. proposed CNN and RNN for PsD detection. J. Zhang vouched for SVM and artificial neural network (ANN) on SPECT scan images. M. Shaban also supported the idea that ANNs are most widely used. L. Di Biase et al. highlighted the shortcomings of the algorithms used in 2020. A. Rana et al. obtained the best accuracy to diagnose PsD for voice data using L1-Norm SVM along with K-fold cross-validation, for handwriting samples, using bagging ensemble, and for gait, it was obtained using SVM. Chandrabhatla A.S. et al. and K. Giannako-poulou explained the evolution of PsD detection techniques and the role IoT could play in it, respectively. Kumar S. et al. proposed accelerometer data crucial for gait analysis. In 2023, S. Dixit et al. had summarized the present trends and the future prospects in detail. However, a very few authors have covered the causes of the disease, information about the datasets, and the classification of patients into different stages of PsD. Also, most of the authors have worked on a few modalities only. Moreover, this is the first survey to our knowledge in which the prospect of EI being used to detect PsD is being discussed.

The survey's primary contributions are:

A novel modalities-based taxonomy for PsD detection is proposed.

The roadmap of the proposed survey is depicted in Figure 2. The present paper is arranged in the following manner: Section 2 presents a descriptive outline of the PsD detection techniques using AI based on various modalities. Section 3 describes the techniques for predicting the different stages in PsD. Section 4 highlights the importance, dimensions, and the assessment measures of EI. Section 5 describes the techniques for prediction of PsD using EI. Section 6 summarizes the various modalities and datasets available for PsD. Section 7 and 8 summarize the research gaps and future scope in this area, respectively. Section 9 sums up the article with the conclusion drawn after reviewing all the previous research and review articles.

Figure 2:

Heena Kathuria et al. utilized 3T venous blood oxygen level-dependent and high-definition susceptibility-weighted imaging sequences in the MRI images to find a potential imaging biomarker in PsD diagnosis [31]. Stephanie Mangesius et al. used a multimodal approach integrating serum and MRI biomarkers to characterize all subtypes of parkinsonism, i.e., multiple system atrophy (MSA), progressive supranu-clear palsy (PSP), and PsD with higher accuracy for all the three diagnoses [56]. Sivaranjini and Sujatha [57] used a pre-trained AlexNet for 2-D MRI image categorization from T2-weighted MRI images from the Parkinson's Progression Markers Initiative (PPMI) collection to enhance PsD diagnosis. The accuracy of the suggested AlexNet architecture was 88.90%, the true positive rate (TPR) of 89.30% and the true negative rate (TNR) of 88.40%, respectively. Solana-Lavalle and Rosas-Romero [58] suggested a model employing ML techniques such as K-nearest neighbors (KNN) and support vector machine (SVM). Voxel-based morphometry was used to extract the regions of interest in MRI images from the PPMI data-set. By training an optimized AlexNet model on digital imaging and communications in medicine (DICOM) images from Istanbul University's hospital, Huseyn [59] created a model for the early identification of PsD. Performance of AlexNet and GoogleNet is contrasted along with the distinction of PsD and MSA. Tarjni Vyas et al. used the flair and T2-weighted MRI scans after applying Z-score normalization from the PPMI repository and 2D and 3D CNN were implemented giving the accuracy of 72.22% and 88.9%, respectively [5]. Sabyasachi Chakraborty et al. used the 3T T1-weighted MRI images from the PPMI database to detect PsD [60]. Cui et al. [61] proposed the use of Resnet18 and SVM to improve classification accuracy by using the fusion of image texture features with deep features, and Sangeetha et al. [62] recommended the use of CNN for the classification of PsD and healthy controls (HCs) using PPMI MRI data. A recently investigated imaging biomarker for the detection of neurodegenerative parkinsonism, including PsD, PSP, and MSA, is the loss of dorsolateral nigral hyperintensity based on MRI data [31]. Table 3 summarizes the learnings from the various recent studies carried out on MRI data, and it is concluded that Cui et al. [61] have achieved the maximum accuracy of 98.66% on the PPMI dataset.

PsD diagnoses are mostly based on the clinical assessments and hence not easily accessible to all. Also, the rate of incorrect diagnoses is substantial; thus, with the advancement in medical imaging technology, automated PsD diagnosis with high accuracy has become significant. Positron emission tomography (PET) is a type of functional imaging method that scans and measures changes in metabolism as well as in physiologic functions such as blood flow, local chemical properties, and absorption. Radioactive chemicals called radiotracers are used in PET. Distinct tracers are utilized for various imaging objectives according to the target process within the body [65]. Dopaminergic and serotonergic PET imaging helps to monitor the progress of motor and nonmotor symptoms and problems as PsD advances. Metabolic, cholinergic, and β-amyloid PET are effective methods for inspecting the cognitive decay in PsD [66]. Dai et al. [67] used PET scans of PsD patients for computer-aided detection. The sole way for the brain to get energy is through glucose uptake; hence, variations in glucose levels in different parts of the brain can indicate the severity of PsD. According to PET scans, PsD patients have a higher glycol consumption in the parietal cortex and a lower one in the frontal brain, which gets worse as time goes on. PET reveals alterations in the affected brain's glucose metabolism, which may lessen the lesion's distribution.

A useful fluorodeoxyglucose (FDG)–PET-based SVM classifier can accurately detect the advancement of dementia in PsD–mild cognitive impairment (MCI) patients with an accuracy of 74.3%, sensitivity of 95%, and specificity of 91% according to Booth et al. [68]. Wu et al. [69] suggested that radiomic features of 18F-FDG PET images serve as an early biomarker for the detection of PsD using SVM producing an accuracy of 90.97% ± 4.66%. Modi et al. [13] used the VGG16-based CNN system and achieved specificity of 97.5%, accuracy of 84.6%, sensitivity of 71.6%, and precision of 96.7% on a PET dataset from PPMI. Sun et al. [70] proposed a novel DL-based radiomics (DLR) model to diagnose PsD based on FDG PET images with an accuracy of 95.17%. It can be concluded from the above literature that the DLR model achieved the maximum accuracy of 95.17%. PET is an invasive, time-consuming, expensive, and uncommon imaging modality used to detect PsD. The major success of PET scans is that it is able to identify cellular level alterations in cells with high efficiency [71].

Diagnosing the PsD-associated dopamine degeneration automatically using SPECT scans is comparable to the diagnosis performed by healthcare experts [44]. A simple, yet fast ANN was recommended by Rumman et al. [72], which was trained on 200 SPECT pictures of brains from the PPMI repository. The model could detect PsD with an accuracy of 94%, sensitivity of 100%, and specificity of 88%. Hathaliya et al. [73] diagnosed PsD and its progression to an accuracy of 88% by training the CNN model on SPECT images. Adams et al. [32] worked on the data of 252 subjects from the PPMI repository and validated that adding DAT SPECT scans and UPDRS_III scores to DL models significantly improves the outcome. Magesh et al. [74] used the CNN and the VGG16-based TL scheme to obtain an accuracy of 95.2% using DaTscan SPECT images. Pereira and Ferreira [75] classified the PsD patients from HC up to an accuracy of 65.7%. Wenzel et al. [76] applied the CNN and ImageNet-based TL, semi-quantitative striatal binding ratio (SBR) analysis to come up to an accuracy of 97%. Ortiz et al. [77] used DaTscan images to detect PsD to an accuracy of 95% using a CNN model. Some of the latest research work done to detect PsD using SPECT images is highlighted in Table 4, specifying the pros and cons of each. H. Khachnaoui et al. [54] reviewed various techniques to diagnose PsD using SPECT scans and concluded that hand-crafted ML algorithms and DL are an active research field. It is observed that Wenzel et al. [76] has achieved the highest accuracy of 97% in identifying PsD on the PPMI dataset. Also, the SPECT scan is found to identify PsD at an early stage [78].

Speech data are used for PsD identification as the speech defects occur at the early stages of the disease [79]. According to Yadav et al. [28], speech deficits in PsD produce symptoms such as decline in the vocal tract volume, diminished intensity of speech and the pitch-range, reduced tongue flexibility, deterioration in quality of speech, irregular pronunciation, and improper halts in speech. The test results show that the DT group effectively evaluated PD and HC with an accuracy of 94.87%. Agarwal et al. [80] suggested that the extreme learning machine (ELM) approach recognizes PsD and HC with an accuracy of 90.7% and a MCC of 0.81. The proposed method achieved 81.5% accuracy when tested against a separate dataset of PsD patients. The SVM classifier was applied by Haq et al. [81] for the accurate classification of PsD patients and HC. For accurate target classification, the L1-norm SVM feature selection was utilized to choose pertinent and closely related characteristics. Based on the value of the feature coefficient weight, it created new feature subsets from the PsD dataset. The best suitable values for the tuning parameters for the classification model were chosen using the k-fold cross-validation method. An accuracy of 99%, sensitivity of 100%, and specificity of 99% were attained on average by the classifier after 10 folds of cross-validation. Z. Soumaya et al. applied the wavelet transform, and the MFCC coefficients are extracted from a dataset of audio recordings of PsD patients and HC while pronouncing the vowel “a.” To accurately classify the sick and healthy people, the KNN classifier is used to give an accuracy of 98.68%. It is observed that larger training data yield more precise results [82]. Karabayir et al. [83] stated that ML can detect PsD accurately using noninvasive and cost-effective speech recordings. Light Gradient Boosting produced 84.1% accuracy as compared to other ML algorithms on UCI–ML datasets comprising 44 speech-test-based acoustic features. Soumaya et al. [84] propose applying the Genetic Algorithm and the classifier SVM on voice recordings obtaining the best accuracy of 91.18%. It minimizes the dimension of the features vector and maximizes the accuracy. Reddy and Alku [79] proposed a non-negative least squares-based exemplar-based sparse representation for the PsD classification approach, which performs better than the traditional ML-based methods by obtaining an accuracy of 82.84% on the PC-GITA PsD dataset and an accuracy of 83.08% on the MDVR-KCL database. From the above literature, it can be concluded that the SVM using the k-fold cross-validation method gives the highest accuracy of 99% on speech recordings.

Rapid eye movement sleep behavior disorder is an often-overlooked sleep disorder harming 70% of the PsD patients. Saccade, pupil, and blink responses could be an early biomarker for PsD detection according to Perkins et al. [30]. Stuart et al. [85] also asserted that monitoring of the saccades may serve as a potential noninvasive early biomarker for long-term PsD cognitive decline and identification. Raschellà et al. [86] investigated the use of wrist actigraphy for easy and quick RBD detection in the comforts of home, on 26 PsD patients achieving an accuracy of 92.9 ± 8.16% using ML classification algorithms. Bek et al. [87] identified that the eye movements disclose subtle repercussions of motion on the emotions processing in PsD. Hence, they compared the realization of static as well as dynamic emotional expressions in PsD using eye-tracking. Accuracy and eye movement measures were analyzed using SPSS. Repeated-measures ANOVAs were used to determine the effects of motion and emotions on the participants. The results imply that the older adults with PsD may use motion differently from healthy older adults when recognizing facial emotions. It can be concluded that the use of AI is still very limited and has good scope for further research in the eye-tracking modality for PsD detection.

Freezing of gait (FoG), i.e., the inability to move, is a typical gait symptom in PsD. Masiala et al. [88] proposed a deep RNN with LSTM cells to detect FoG in PsD with an AUC score of 93%, specificity of 85%, and sensitivity of 89%. A vision-based PsD detection system can be used to assess the presence of PsD using a feature-weighted minimum distance classifier model. Researchers have applied a variety of technologies such as accelerometers, integrated load sensors in shoes, gyroscopes, inertial measurement units, pressure-sensitive walkways or other laboratory setups, and video data collected by camera, smart-phone, or any device in-built with video-recording facility to objectively capture gait parameters. Rupprechter et al. [89] used an ordinal Random Forest classifier that was trained to estimate UPDRS severity scores. It achieved high performance. Computer vision-based modality involves two subtypes, i.e., marker-based, and marker-less. In marker-based, the human body model is constructed manually using sensors to extract relevant features, whereas in marker-less, i.e., the appearance-based modality is that computer vision techniques are used to extract the gait parameters [90, 91]. Kour et al. [90] revealed that the SVM classifier is analyzed to be the most popular among researchers to effectively classify PsD and HC. Using six-angle features in a marker-based motion capturing system for person recognition, Zahra et al. [91] obtained an accuracy of 99.3% using KNN. In the smart healthcare setting, the gait feature selection from the recorded skeletal points is employed to help the real-time PsD diagnosis. In comparison with the current healthcare systems, Sathya Bama and Bevish Jinila [92] asserted that SVM and ensemble learning-based classification models offer faster prediction and accurate classification. The regions of interest (ROIs) are segmented using an enhanced vision-based approach, and gait parameters—namely, spatiotemporal (SPT), linear kinematic and other body motion features—are assessed using statistical tools. Kour et al. [93] have categorized PsD and HC using ML algorithms such as SVM, KNN, Random Forest, and linear regression. KNN outperformed the other models at three severity levels with an accuracy of 90.06%. Pei et al. [94] presented a temporal PAST fusion model, which enhanced the classification accuracy for both PsD and HC classification and PsD level classification on the PhysioBank dataset with an accuracy of 99.18%. Xia et al. [95] proposed a dual-modal model, with each branch having a convolutional network followed by an attention-enhanced bi-directional LSTM; referred to as DCALSTM. It produced an accuracy of 99.31%. We can summarize that the DCALSTM and the KNN achieved the highest accuracies. Also, the motion capture systems are popular and act as the golden-standard methodology for clinical gait analysis, but pervasive gait analysis in home-based environments using wearable systems including pressure insoles, inertial sensors, and EMG sensors is required [96].

Bhurane et al. [97] used a time-domain technique for the detection of PsD using EEG. An SVM classifier with the third-degree polynomial kernel is used to diagnose PsD automatically utilizing the interchannel similarity features, correlation coefficients (CCs), and linear predictive coefficients (LPCs). Selected characteristics produced using the feature ranking and principal component analysis (PCA) approaches are used in a progressive feature addition manner. A maximum accuracy of 99.1% is attained. Chawla et al. [98] employed a flexible analytic wavelet transform (FAWT). The EEG signals are first preprocessed and then divided in five frequency sub-bands. Through the analysis of variance (ANOVA), a number of entropy parameters are calculated from the decomposed sub-bands and rated according to their significant level in identifying PsD. Classifiers such as SVM, logistics, radial basis functions (RBFs), Random Forests (RFs), and KNN are used to determine the appropriate feature sets. KNN gives the best accuracy of 99% on Dataset1 of EEG recordings of 20 PsD and HC, respectively. Wagh and Varatharajah [99] proposed an EEG–GCNN model, which substantially surpasses the human and the ML baselines, with an AUC of 0.90. It captured both the spatial and functional associativity between the scalp electrodes. Coelho et al. [100] identified PsD patients using the Hjorth features extracted from the EEG signals. These features displayed a major difference between the HC and PsD, especially when computed in the frontal, central, parietal, and occipital zones. Supervised ML algorithms, KNN and SVM were used, and SVM showed higher accuracy of 89.56% in the classification. Khare et al. [101] uses EEG signals to present an accurate and robust PsD detection model using smoothed pseudo-Wigner Ville distribution coupled with convolutional neural networks producing the best PsD detection accuracy of 100% and 99.97% for datasets 1 and 2, respectively. But, the shortcoming in this model is that the datasets used comprise very few subjects. It can be observed from the following literature that a few studies have managed to achieve the high accuracy of 99% and above, but it needs to be validated using larger datasets.

None of the techniques to diagnose the PsD at an early stage is completely effective. Hence, extensive research is being carried out to identify newer modalities and improved techniques. An optimized form of the crow search algorithm (OCSA) is therefore introduced by Gupta et al. [102] to detect PsD. The OCSA may be used to accurately predict PsD and assist people in receiving the right care at an early stage. The effectiveness of OCSA's performance on 20 benchmark datasets is commendable, and the outcomes have been contrasted with that of the chaotic crow search algorithm (CCSA). The suggested approach is more stable and discovers an ideal subset of characteristics while optimizing accuracy to 100% and lowering the number of features chosen. Nasser et al. [103] applied noise removal and data augmentation by rotation, flipping, contour and freeze and fine-tuning on the PaHaW dataset and achieved an accuracy of 98.28% using AlexNet classifier with transfer learning technique and they vouch that the spiral dataset is more efficient for Parkinson's detection. Gazda et al. [29] worked on the PaHaW and HandPD dataset and by applying label-preserving and optimal image transformations and skeletonization, they achieved an accuracy of 94.7% using the CNN. The highest accuracy of 100% is achieved by D. Gupta et al., but it is not validated for the various UDPRS stages.

Videos and images of both healthy and PsD patient faces are used as input by Face++. Facial expression characteristics (the amplitude of facial expression and shaking of tiny facial muscle fibers) are extracted using relative positions and positional jitter. When the placements of the essential features are applied, an F1 value of 99% can be attained by using an SVM model for the diagnosis of PsD. Doctors may be able to undertake remote diagnostics using the information to better grasp the disease's current dynamics [104]. Sonawane and Sharma [105] observed that hypomimia is a significant biomarker of PsD and the CNN can diagnose PsD with 85% accuracy. Mattavelli et al. [184] achieved an accuracy of 83.34% by using the SPSS statistical software, who asserted the findings that the facial expression recognition, particularly the fear emotion, is critically impaired by neurodegeneration in PsD and that it appears before other cognitive impairments. It is observed that efficient training of facial expression recognition models requires large-scale training data, hardware advancements, such as GPUs, and lots of high-quality data. Presently, there is much scope of research in this field [106].

The comparative analyses of the highest accuracies achieved in various modalities in PsD detection are shown in Figure 4.

Figure 4:

MRI discloses certain anomalies caused due to the buildup of iron in the brain; hence, it can be used to identify the biomarkers and verify whether the patient suffers from PsD or not. Bhattacharya et al. [187] used a CNN, taking the MRI images of the brain as input, to understand the latent features layer by layer and eventually identifying the required biomarkers for the classification [5]. An incremental SVM has also been used to identify the total-UPDRS and the motor-UPDRS scores. Data dimensionality reduction is accomplished using nonlinear iterative partial least squares, while clustering is accomplished using self-organizing maps. With an accuracy of 94%, CNN can identify PsD and stage it into distinct categories, according to Chakrabarti et al. [108]. They applied min–max normalization on the PPMI database and used a CNN model with five convolutional layers with K-cross field validation to diminish bias.

Research has been conducted on the progression of PsD using DL and ML techniques on the telemonitoring dataset. Kathuria et al. [31] asserted that the classification based on the motor UPDRS score performs better than the classification on the total UPDRS score and thus it can be established as a better measure for severity prediction. The voice data of PsD patients are collected for analysis and normalized into 16 biomedical voice measures using min–max normalization. The DNN identified the two classes—“severe” and “non-severe.” Speech defects in PsD patients are mostly caused by the lowered lung capacity, impaired facial muscle function, impaired laryngeal function, and diminished speaking drive. Many abnormalities in voice and speech result from these changes, including decreased volume, difficulty with volume changes, limited voice modulation (monotonous speech), decreased vocal fold tension, hoarse tone, insufficient articulation leading to slurred speech, and change in speech pace. The condition is known as hypokinetic dysarthria. Phonation, articulation, and prosody defects in speech are caused by injury to the neural pathways and centers that control the nerve impulses of the speech organs. Reduced amplitude and motion tempo of the lips, jaws, and tongue induce alterations in articulation. Reduced accentuation and incorrect consonant articulation lead to babbling. Identification of acoustic features was done using SVR, MLR, and RF. The best performing algorithm was Random Forest Regressor [56]. Sajal et al. [109] combined the tremor and voice data analysis for diagnosing PsD. Wavelet filter banks preserved the tremor signals' time-localized transient nature and different UPDRS levels were distinguished. KNN outperforms the other models in terms of accuracy, specificity, and sensitivity. Raza et al. [110] used XGBoost on speech signals for monitoring the progression of PsD. Nilashi et al. [111] ascertained that the hybrid of clustering and deep belief network (DBN) with the aid of support vector regression (SVR) for an ensemble of the DBN outputs is capable of relatively better detection of total-UPDRS and motor-UPDRS scores than other learning techniques.

Any characteristic used to measure human intelligence must adhere to the rules of psychometrics, the study of psychological assessment. The EI assessment must show that it is capable of more than just capturing data on personal attributes or other capabilities. Three techniques are frequently employed to evaluate EI:

PsD patients apprehend arousal better over valence, and among emotional categories, disgust, fear, and surprise less accurately, while sadness is most accurately identified. Near-perfect PsD vs HC identification is achieved via emotional EEG responses. Overall, the study shows that (a) analyzing implicit reactions alone allows (i) the discovery of valence-related deficits in PsD patients and (ii) the differentiation of PsD from HC and (b) emotional EEG analysis is an ecologically legitimate, efficient, simple, and long-lasting tool for PsD detection in comparison to expert assessments, self-reports, and resting-state analysis [124]. Another study found that while PsD did not demonstrate any behavioral abnormalities, it did display deficiencies in the processing of emotional data as measured by neurophysiological tests. This demonstrated the potential for distributed spectral power in various frequency bands to offer insightful data on emotional processes in PsD patients. When compared to HC, PsD patients showed lower overall relative delta, theta, alpha, and beta power as well as lower alpha, absolute theta, and beta power at the bilateral anterior regions as well as greater mean total spectrum frequency in EEG signals over various emotional states. Differences in alpha, theta, and beta power asymmetry indices between the right and left hemispheres were observed in the controls. Patients showed bilateral intrahemispheric alpha power imbalance decrease across all regions. During emotional events, discriminant analysis accurately identified 95.0% of the patients and controls [125].

The CNN is used for PsD diagnosis from facial expressions. The CNN includes convolution layers for identifying features and fully connected layers for PsD detection. The diagnosis capability of various facial emotions obtained the most effective output with an accuracy of 96.5% and an AUC of 95.3%. This technique is a noninvasive and reliable method for the diagnosis of PsD [126].

PsD can be detected by the analysis of emotion changes during pronunciation-defined speech exercises. Using the XGBoost, a balanced accuracy of 69% was attained on the speech data of a Czech tongue twister. It can be employed in clinical practice because the features are explicable. In this research, fear has been found to be the most effective emotion for PsD identification. From a statistical perspective, the characteristics reflecting variations in how people expressed anxiety at the time were extremely important for the PsD diagnosis [22].

Figure 5 presents the process flow of PsD detection using EI. Step 1, the datasets can be taken from a secondary or primary source. Preprocessing techniques such as missing values imputation, normalization, etc. are applied to the raw data. An appropriate model is selected that performs the task of PsD vs HC classification while improving the accuracy, specificity, and sensitivity performance metrics. A comparative analysis of PsD detection using EI is presented in Table 6. It is observed that Parameshwara R. et al. obtained the highest accuracy of 99% using KNN and Naïve Bayes Classifiers on EEG signals using EI.

Figure 5:

Speech, handwriting, gait, and imaging modalities such as MRI, PET, SPECT, eyes, facial expression recognition, EEG signals, etc., are the various modalities for PsD detection. Figure 6 defines the taxonomy of PsD detection based on the most widely used modalities.

Figure 6:

Datasets have an essential and foundational function in AI-based learning and diagnosis systems. Various forms of data inputs for distinguishing the PsD patients from HC are reported in the literature. Over the years, a variety of datasets have been compiled and made accessible to professionals and aspiring AI and ML researchers for the purposes of analysis and experimentation. This section provides a comprehensive overview of the public and private PsD datasets available for each of these dataset categories.

Several research studies have been carried out for the collection of Parkinson's datasets based on various modalities. Publicly available datasets promote research extensively as the data are free to use for all and are presented in Table 7.

Private datasets fill a major void in the research datasets by offering desirable properties in terms of modalities, constraints, and quality at the same time. Table 8 presents all the private datasets in detail.