A classification prediction method for denoised MRI images of ADHD

Abstract

The present application relates to the technical field of bioinformatics and deep learning, and discloses a kind of classification prediction methods for the denoising MRI image of ADHD, including the MRI image of brain to be detected is collected, the MRI image collected is carried out online denoising processing using the unsupervised denoising network of online training, then the result of online denoising processing is preprocessed again and input into the DSC-Former model of offline training to obtain classification prediction result.Unsupervised denoising network includes two parallel processing lines, each line includes a series of down-sampling module and noise estimation model;DSC-Former model includes four layers of three-dimensional snake convolution module and Vision Transformer module connected in turn, and a SoftMax layer.The present application adopts the method of denoising processing first and then classification prediction, which effectively improves the classification prediction effect of MRI image of ADHD patient.

Description

Technical Field

[0001] The present invention belongs to the field of bioinformatics and deep learning technology, and specifically relates to a classification prediction method for denoised MRI images of ADHD. Background Art

[0002] Attention Deficit Hyperactivity Disorder (ADHD) is a common neurodevelopmental disorder affecting children and adults worldwide. Its main characteristics include inattention, hyperactivity, and impulsive behavior. The accurate diagnosis of ADHD relies primarily on clinical symptom observation, behavioral scale assessment, and detailed medical history collection, lacking direct biological markers for confirmation. This subjective diagnostic approach often leads to high misdiagnosis rates, compromising treatment outcomes and patient management.

[0003] At the same time, advances in medical imaging technology, particularly the widespread application of magnetic resonance imaging (MRI) in neuroscience, have provided new insights into the neurobiological underpinnings of ADHD. MRI can non-invasively reveal changes in brain structure and function. Recent studies have revealed differences in brain volume, white matter integrity, and functional connectivity in specific brain regions among individuals with ADHD. However, MRI images are susceptible to noise from various factors, such as motion artifacts and hardware limitations. These factors often compromise the quality of the imaging data and the accuracy of subsequent analysis.

[0004] Deep learning, a revolutionary technology in the field of artificial intelligence, has been successfully applied to medical image analysis with its powerful data processing and feature learning capabilities, significantly improving the accuracy of disease diagnosis. By constructing a multi-level neural network model, deep learning automatically extracts subtle features related to the disease from complex image data, achieving accurate classification of the disease state. For example, the invention application with publication number CN118588288A, "Assisted Screening for ADHD in Children Based on Multimodal Appearance Features," has a data acquisition module for collecting video stream data through a camera; a data preprocessing module for framing the collected video files and obtaining facial regions of interest; a feature extraction module that uses computer vision technology to analyze head posture, expression, and gaze distribution, and performs homogeneous vector fusion; and a multimodal information classification module that uses a temporal convolutional network (TCN) to pre-train the fusion vectors and detection results within a certain time period, and finally determines the classification result of whether the patient has abnormal behavior during the entire time period. However, existing deep learning methods mostly focus on achieving accurate ADHD classification prediction by adopting improved feature extraction networks and target detection methods. Few methods first use denoising methods to improve the image quality of the input image to further improve the classification effect. Therefore, existing technologies need to be improved. Summary of the Invention

[0005] The technical problem to be solved by the present invention is to provide a classification prediction method for denoised MRI images of ADHD, which is used to combine MRI imaging technology with deep learning algorithms for classification prediction of MRI images of ADHD.

[0006] In order to solve the above technical problems, the present invention provides a classification prediction method for denoised MRI images for ADHD, comprising: acquiring an MRI image of the brain to be tested, performing online denoising on the acquired MRI image using an online trained unsupervised denoising network, and then preprocessing the result of the online denoising process and inputting it into an offline trained DSC-Former model to obtain a classification prediction result;

[0007] The unsupervised denoising network includes two parallel processing circuits, each of which includes a downsampling module and a noise estimation model connected in series;

[0008] The DSC-Former model consists of four layers of sequentially connected three-dimensional snake convolution modules and VisionTransformer modules, as well as a SoftMax layer.

[0009] As an improvement of the classification prediction method of denoised MRI images for ADHD according to the present invention:

[0010] The downsampling module is a three-dimensional convolution kernel structure;

[0011] The noise estimation model in the unsupervised denoising network is composed of a stack of Vision Transformer and ConvFormer.

[0012] As a further improvement of the classification prediction method of denoised MRI images for ADHD of the present invention:

[0013] The online denoising process of the unsupervised denoising network is:

[0014] Each MRI image was trained independently for 50 epochs using the Adam optimizer. During the training process, a loss function was used to approximate the denoised image obtained by the noise estimation model to the other downsampled image through backpropagation of the two downsampled images, thereby obtaining noise estimates for the two downsampled images. After training, the two output noise estimates were normalized, and the difference between the original input MRI image and the normalized result was used as the result of the online denoising process:

[0015]

[0016] Among them, θ is the noise estimation model, x is the input image, S1(x) and S2(x) are two downsampled images, and f θ (S(x)) is the two-way noise estimate output by the noise estimation model.

[0017] As a further improvement of the classification prediction method of denoised MRI images for ADHD of the present invention:

[0018] The unsupervised denoising network loss function is:

[0019]

[0020] As a further improvement of the classification prediction method of denoised MRI images for ADHD of the present invention:

[0021] The pre-processing is:

[0022] (1) Mapping the denoised MRI image data to a voxel size of (1,1,1);

[0023] (2) Normalize the intensity value of each voxel and map it to the range of 0-511;

[0024] (3) Crop the non-brain background.

[0025] As a further improvement of the classification prediction method of denoised MRI images for ADHD of the present invention:

[0026] The three-dimensional dynamic snake convolution module in the DSC-Former model is characterized by:

[0027] The module contains four branch paths, using conventional three-dimensional convolution, dynamic snake convolution in the x-direction, dynamic snake convolution in the y-direction, and dynamic snake convolution in the z-direction. The three dynamic snake convolution modules form receptive fields along the X, Y, and Z axes respectively. The feature vectors extracted by the four different convolutional layers are concat-operated to output the corresponding feature representation.

[0028] As a further improvement of the classification prediction method of denoised MRI images for ADHD of the present invention:

[0029] The offline training process of the DSC-Former model is as follows:

[0030] MRI image data of desensitized ADHD patients and a control group were obtained from the public ADHD-200 dataset and cooperative hospitals. Each MRI image was then trained online using the unsupervised denoising network to obtain a denoised MRI image. Preprocessing and data augmentation were then performed, and the augmented dataset was randomly divided into a training set, a validation set, and a test set in an 8:1:1 ratio.

[0031] The MRI images in the training set were sequentially input into the DSC-Former model. DSC-Former was trained using the RMSprop optimizer, and a loss function was constructed based on cross entropy to calculate the loss between the recognition result and the label. After the backward propagation process was completed, the DSC-Former model weights were updated. The classification performance of the model was validated using a 10-fold cross-validation method.

[0032] The beneficial effects of the present invention are mainly reflected in:

[0033] 1. The present invention adopts a method of first denoising and then classification prediction, which effectively improves the classification prediction effect of MRI images of ADHD patients.

[0034] 2. To address the issue of inherent noise sources such as hardware noise affecting image quality during MRI image acquisition, this paper designs an unsupervised MRI image denoising network module. This module uses a three-dimensional downsampling module to perform unsupervised denoising on independent MRI image data. This module is applicable to MRI images of any data size and can reduce the interference of image noise on feature extraction during subsequent classification network recognition, thereby improving classification accuracy. Furthermore, the denoised image edges are clearer, facilitating further observation and judgment of brain structured information by doctors.

[0035] 3. This paper designs a Transformer-based DSC-Former classification network, combining convolution and self-attention to construct a feature extraction network. Furthermore, targeting the characteristics of MRI images, a three-dimensional dynamic snake convolution is used as the network downsampling module. This enables the model to dynamically adjust the convolution kernel sampling position based on input features, making it more suitable for brain MRI image data and further reducing background interference on feature extraction. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] The specific embodiments of the present invention are further described in detail below with reference to the accompanying drawings.

[0037] Figure 1 This is the network structure diagram of the DSC-Former proposed in the present invention;

[0038] Figure 2 This is the structure diagram of the MRI image denoising network downsampling 3D convolution kernel proposed in the present invention;

[0039] Figure 3 This is the structural diagram of the MRI image denoising network proposed in the present invention;

[0040] Figure 4 Schematic diagram of the Transformer and Convformer structures;

[0041] Figure 5 Schematic diagram of the three-dimensional dynamic snake convolution module. DETAILED DESCRIPTION

[0042] The present invention is further described below with reference to specific embodiments, but the protection scope of the present invention is not limited thereto:

[0043] Example 1: A classification prediction method for denoised MRI images of ADHD, such as Figure 1-5 As shown in the figure, the unsupervised MRI image denoising network module is used to perform unsupervised denoising on the MRI images of ADHD, reducing the interference of image noise on feature extraction and improving classification accuracy. The denoised MRI images are then classified and predicted based on the Transformer-based DSC-Former classification network. Specifically:

[0044] Step 1: Data collection and dataset construction for ADHD classification task

[0045] Data collection was done through two main approaches: first, data from the publicly available ADHD-200 dataset; and second, MRI imaging data from desensitized ADHD patients and controls were obtained through cooperative hospitals in Zhejiang Province. Specific dataset information is shown in Table 1 below:

[0046] Table 1. Dataset for ADHD classification task

[0047] Dataset Number of patient groups Number of control groups ADHD-200PKU 113 146 ADHD-200NYU 123 99 ADHD-200OHSU 37 42 ADHD-200Brown 24 27 Data provided by partner hospitals 63 232

[0048] Step 2: Build an unsupervised denoising network

[0049] The unsupervised MRI denoising network structure proposed in this invention is as follows Figure 2 As shown, it includes two parallel processing lines, each of which includes a downsampling module and two noise estimation models consisting of a corresponding Vision Transformer and ConvFormer stack. After the MRI image passes through the two downsampling modules, two downsampled images are obtained, that is, two sampling results on an image with identical noise distribution are obtained. These two downsampled images are respectively used as inputs of a noise estimation model. During the training process, noise estimation is performed on two paired downsampled images of a single MRI image, so that the denoising result of the downsampled image is as close as possible to the other downsampled image, thereby obtaining a noise estimation of the original image.

[0050] Downsampling module such as Figure 3 As shown in the figure, a 3D convolution kernel structure with a convolution kernel size of (2,2,2) is used to perform downsampling by averaging the 3D diagonal voxels. Each MRI image is independently used as input data for the denoising network. After passing through two downsampling convolution kernels, two independent and identically distributed downsampled images of the same image are obtained.

[0051] The noise estimation model is composed of Vision Transformer and ConvFormer stacked together. The specific model structures of Vision Transformer and ConvFormer are as follows: Figure 4 As shown in the figure, in the Vision Transformer, input data passes through the Multi-Head Attention layer and the feed-forward network in sequence. Both the Multi-Head Attention layer and the feed-forward network use residual connections and are normalized using the LayerNorm layer. Compared to the Vision Transformer, the ConvFormer module only replaces the Multi-Head Attention layer with the SeparableConvolution layer, while the rest of the components remain the same.

[0052] After the downsampling module, two downsampled images are obtained, which are used as the input of two noise estimation modules respectively. The noise estimation results of the downsampled images are normalized, and then the difference between the original input MRI image and the normalized result is used as the final denoised image, thereby realizing the noise estimation of the original image and finally obtaining the denoised image.

[0053]

[0054] Among them, θ is the noise estimation model, x is the input image, S1(x) and S2(x) are two downsampled images, and f θ1 (S(x)) and f θ2 (S(x)) are the two noise estimates output by the noise estimation model.

[0055] The denoising network loss function for model training is designed as follows:

[0056]

[0057] Through this loss function, the denoised image of S1(y) is made as close as possible to S2(y), and the denoised image of S2(y) is made as close as possible to S1(y). Thus, after training, the noise estimation of the two downsampled images is obtained. Step 3: DSC-Former model construction

[0058] The DSC-Former model proposed by the present invention is as follows Figure 1 As shown in the figure, the backbone network of the DSC-Former model consists of a three-dimensional dynamic snake convolution module and a Vision Transformer module. The three-dimensional dynamic snake convolution module integrates directional snake convolutions in the X, Y, and Z axes, focusing on extracting slender structures in different directions to adapt to the characteristics of brain MRI images used in ADHD diagnosis. This module uses Leaky RELU as the activation function; the Vision Transformer module uses StarReLU as the model activation function, which can achieve better performance than GELU with less computation. The model inference path passes through the snake convolution module and the Vision Transformer module in sequence, and the final output layer uses a SoftMax layer.

[0059] The module structure of Dynamic Snake Convolution (DSC) is as follows: Figure 5 As shown in Figure 2, as a downsampling model, each layer extracts feature information from MRI images. DSC is inspired by deformable convolution and introduces a dynamic offset to adjust the position of the convolution kernel. Unlike deformable convolution, DSC introduces a prediction network in three dimensions to predict the dynamic offset of each pixel, and introduces a smoothness term to constrain the change of the offset, so that the convolution kernel can move smoothly along the direction of the tubular structure, like a snake, thereby better adapting to slender structures and having good feature extraction capabilities for complex brain tissue scenes. Taking the x direction as an example, the calculation process of its receptive field position is shown in formula (3), (xi+c y i+c z i+c ) is the center of the convolution kernel, thus forming a snake-like receptive field along the x direction:

[0060]

[0061] The three-dimensional dynamic snake convolution module used in the present application integrates conventional three-dimensional (3D) convolution and three directional dynamic snake convolution—X, Y and Z. The conventional 3D convolution is responsible for extracting the key features of the target object in the three-dimensional image. The X, Y and Z dynamic snake convolution focuses on extracting the characteristics of the elongated tubular structure along the X, Y and Z axis direction respectively. Finally, the feature vectors extracted from the four different convolution layers are concatenated together to form a comprehensive feature representation. As shown in Figure 5 The DSC module is divided into four branches, where Conv3D is a conventional three-dimensional convolution kernel, and the DSConv3D module in the other three branches is used to learn the feature information in three directions. DSConv3D is connected by two Conv3D modules in residual, and the Instance Norm layer is set after the residual part, and the Tanh activation function is used. The Instance Norm layer is set after each convolution kernel, and the Leaky ReLU activation function is used.

[0062] Between each layer of the DSC module, the Vision Transformer is used to further learn the feature information in a self-attention manner. The output of the three-dimensional dynamic snake convolution module is added to the Instant Norm layer, and the Leaky RELU is used as the activation function.

[0063] The Vision Transformer extracts the correlation information between regions of image data through the self-attention mechanism, thereby extracting the image markers related to ADHD contained therein. At the same time, the Vision Transformer module used in the present application uses StarReLU as the activation function. The StarReLU calculation process is as follows:

[0064] StarReLU(x)=s·(ReLU(x)) 2 +β (4)

[0065] By introducing two trainable parameters s and β, better performance than GELU can be achieved with less computational cost.

[0066] The model inference path successively passes through four snake convolution modules and Vision Transformer modules, and finally passes through the SoftMax layer to obtain the classification prediction result given by the model.

[0067] Step 4: Denoise, preprocess and normalize the MRI image

[0068] The unsupervised denoising network constructed in step 2 is used to denoise all MRI images in the dataset constructed in step 1. The denoising process for each MRI image is a separate training process. During the training process, the loss function of formula (2) is used to make the two downsampled images pass through backpropagation, and the denoised image obtained by the noise estimation model is as close as possible to the other downsampled image, thereby obtaining noise estimates for the two downsampled images. After the training is completed, the two output noise estimates are normalized, and then the difference between the original input MRI image and the normalized result is used as the denoising result of the unsupervised denoising network.

[0069] Then all MRI images after denoising are preprocessed and normalized, which mainly includes the following steps:

[0070] (1) All image data are resampled and mapped to a voxel size of (1,1,1).

[0071] (2) Normalize the intensity value of each voxel and map it to the range of 0-511.

[0072] (3) Integrate the training data, count the foreground and background information of the image, and crop the non-brain background to unify the image dimensions. The final input data dimension is (128, 128, 128).

[0073] Step 5: DSC-Former model training

[0074] The training process is as follows:

[0075] (1) Each sample in the dataset that has undergone denoising and preprocessing in step 4 is randomly horizontally flipped and rotated to expand the dataset size. The expanded dataset is randomly divided into a training set, a validation set, and a test set in a ratio of 8:1:1.

[0076] (2) Build the DSC-Former model, randomly initialize the model parameters and load the divided training set and validation set. Input the data in the training set into the DSC-Former model in batches to obtain the inference result, calculate the loss function with the input label, complete the back propagation process, and update the DSC-Former model weight. Cross entropy is used to build the binary classification task loss function during training, and the optimizer is set as follows: RMSprop optimizer is used, the learning rate lr is set to 1e-4, the decay rate p is set to 0.9, the weight decay is set to 1e-4, and the ∈ is set to 1e-8. The training Batch size is set to 4, and a total of 200 epochs are trained. The model performance is evaluated on the validation set every round, and the model weight under the highest AUC is reserved. The test set is used to verify the classification AUC, accuracy and recall rate of the optimal weight model.

[0077] (3) 10-fold cross-validation is used in the experiment, and the data set is randomly divided into 10 parts, each part of data is used as a test set in turn, and the remaining data is divided into a validation set and a test set in a 1:8 ratio. Repeat step (2) training process. Finally, the average classification index under 10-fold cross-validation is obtained.

[0078] Step 6, online use

[0079] After the model is trained, the online inference uses the following process:

[0080] (1) Collect the brain MRI image of the patient to be detected;

[0081] (2) Use the unsupervised denoising network proposed in the application for online training and denoising processing of the collected MRI image. Each MRI image denoising process is a separate training process. Each MRI image is independently trained for 50 epochs, that is, the noise estimation process is completed. The loss function used is shown in formula (1), and the Adam optimizer is used for training. After training, the noise estimation of the two down-sampled images is obtained respectively. The noise estimation results of the down-sampled images are homogenized, and then the difference between the original input MRI image and the homogenization result is used as the noise estimation of the original image. Thus, the denoising is completed on the basis of the original image, as shown in formula (1);

[0082] (3) Preprocess the denoised image, resample all image data to (1, 1, 1) voxel size, normalize the intensity value of each voxel, map it to the range interval of 0-511, and crop the non-brain background;

[0083] (4) Input the preprocessed image into the DSC-Former model trained in step 5, and output the classification result after passing through the multi-layer dynamic snake convolution module and Vision Transformer.

[0084] experiment:

[0085] 1. Verification of the effectiveness of unsupervised denoising network

[0086] In order to verify the role of the unsupervised denoising network proposed in the present invention in the ADHD classification and recognition task, an ablation experiment for the denoising network was designed. In the first case (denoised data), the MRI images in all data sets were trained with the DSC-Former model after the unsupervised denoising network and preprocessing process of the present invention. In the second case (original data), the original MRI images were used directly, and the DSC-Former model was trained after the preprocessing operation of the present invention. Under 10-fold cross validation, the ADHD classification indicators of the DSC-Former model in both denoised and undenoised data are shown in Table 2 below:

[0087] Table 2

[0088] method AUC Accuracy (%) Recall rate (%) (Raw Data)DSC-Former 0.921 91.3 84.4 (Denoised Data)DSC-Former 0.948 93.4 89.7

[0089] From the above results, it can be seen that by removing the inherent noise in the MRI image through the unsupervised denoising network of the present invention, the DSC-Former model of the present invention can more accurately extract the feature information in the image and obtain better recognition effect.

[0090] Experiment 2: Comparative experiment

[0091] To further validate the proposed method's ability to classify and identify ADHD patients, this section compared the unsupervised denoising network and DSC-Former model presented in this paper with ADHD classification models proposed in two other related papers on a dataset consisting of the ADHD-200 dataset and data provided by partner hospitals. Reference 1 constructed four different convolutional neural network architectures, of which the ResNet, which performed best, was listed here. Reference 2 used RBSVM to classify ADHD on structural MRI images after feature extraction.

[0092] The training dataset is also divided into two cases. In the first case (denoised data), all MRI images in the dataset are subjected to the unsupervised denoising network and preprocessing process of the present invention before model training. In the second case (raw data), the original MRI images are directly used and subjected to the preprocessing process of the present invention before model training.

[0093] The classification performance comparison experimental results based on the dataset used in this experiment under 10-fold cross validation are shown in Table 3 below:

[0094] Table 3

[0095]

[0096]

[0097] Document 1 See Cicek, Gulay, and Aydin Akan. "Deep learning approach versus traditional machine learning for ADHD classification." 2021MedicalTechnologies Congress (TIPTEKNO). IEEE, 2021.

[0098] Reference 2: Lohani, Dhruv Chandra, and Bharti Rana. "ADHD diagnosis using structural brain MRI and personal characteristic data with machine learning framework." Psychiatry Research: Neuroimaging 334 (2023): 111689.

[0099] The results show that the DSC-Former model of the present invention has an accuracy improvement of 4.8% and 10.2% respectively on the same data set compared to the recognition methods used in References 1 and 2, and has achieved the highest AUC. It can be seen that DSC-Former, by fusing three-dimensional dynamic snake convolution with the Vision Transformer multi-head self-attention mechanism, has shown stronger representation capabilities and recognition accuracy in the ADHD classification task using structural MRI data. When using denoised data, both DSC-Former and the ResNet used in Reference 1 showed better classification performance. It can be seen that the unsupervised denoising network used in the present invention, after removing the inherent noise in the MRI image, enables the classification model to better extract specific information from the image and achieve higher classification accuracy. Since the model used in Reference 2 is based on the brain gray matter and cortical thickness features extracted from MRI images, it is not sensitive to denoised image data and the classification accuracy does not show a significant improvement.

[0100] Finally, it should be noted that the above examples are merely specific embodiments of the present invention. Obviously, the present invention is not limited to the above examples and is subject to numerous variations. All variations that can be directly derived or conceived by a person of ordinary skill in the art from the disclosure of the present invention are considered to be within the scope of protection of the present invention.

Claims

1. A classification prediction method for denoised MRI images of ADHD, characterized by: The method involves collecting MRI images of the brain to be tested, performing online denoising on the collected MRI images using an online trained unsupervised denoising network, and then preprocessing the online denoising results before inputting them into an offline trained DSC-Former model to obtain classification prediction results. The unsupervised denoising network includes two parallel processing circuits, each of which includes a downsampling module and a noise estimation model connected in series; The DSC-Former model consists of four layers of sequentially connected three-dimensional snake convolution modules and Vision Transformer modules, as well as a SoftMax layer; The downsampling module is a three-dimensional convolution kernel structure; The noise estimation model in the unsupervised denoising network is composed of a stack of Vision Transformer and ConvFormer; The online denoising process of the unsupervised denoising network is: Each MRI image was trained independently for 50 epochs using the Adam optimizer. During the training process, a loss function was used to approximate the denoised image obtained by the noise estimation model to the other downsampled image through backpropagation of the two downsampled images, thereby obtaining noise estimates for the two downsampled images. After training, the two output noise estimates were normalized, and the difference between the original input MRI image and the normalized result was used as the result of the online denoising process: Among them, θ is the noise estimation model, x is the input image, S1(x) and S2(x) are two downsampled images, and f θ (S(x)) is the two-way noise estimate output by the noise estimation model; The loss function of the unsupervised denoising network is:

2. The classification prediction method for denoised MRI images for ADHD according to claim 1, characterized in that: The pre-processing is: (1) Mapping the denoised MRI image data to a voxel size of (1,1,1); (2) Normalize the intensity value of each voxel and map it to the range of 0-511; (3) Crop the non-brain background.

3. The classification prediction method for denoised MRI images for ADHD according to claim 2, characterized in that: The three-dimensional dynamic serpentine convolution module in the DSC-Former model is characterized by: The module contains four branch paths, using conventional three-dimensional convolution, dynamic snake convolution in the x-direction, dynamic snake convolution in the y-direction, and dynamic snake convolution in the z-direction. The three dynamic snake convolution modules form receptive fields along the X, Y, and Z axes respectively. The feature vectors extracted by the four different convolutional layers are concat-operated to output the corresponding feature representation.

4. The classification prediction method for denoised MRI images for ADHD according to claim 3, characterized in that: The offline training process of the DSC-Former model is as follows: MRI image data of desensitized ADHD patients and a control group were obtained from the public ADHD-200 dataset and cooperative hospitals. Each MRI image was then trained online using the unsupervised denoising network to obtain a denoised MRI image. Preprocessing and data augmentation were then performed, and the augmented dataset was randomly divided into a training set, a validation set, and a test set in an 8:1:1 ratio. The MRI images in the training set were sequentially input into the DSC-Former model. DSC-Former was trained using the RMSprop optimizer, and a loss function was constructed based on cross entropy to calculate the loss between the recognition result and the label. After the backward propagation process was completed, the DSC-Former model weights were updated. The classification performance of the model was validated using a 10-fold cross-validation method.

Images