The dataset is downloaded from the Harvard Medical School website (Johnson and Becker,
Edited by: Robertas Damasevicius, Silesian University of Technology, Poland
Reviewed by: Afshin Shoeibi, K.N.Toosi University of Technology, Iran; Delaram Sadeghi, Islamic Azad University of Mashhad, Iran; Ritys Moskoliunas, Vytautas Magnus University, Lithuania
†These authors have contributed equally to this work
This is an open-access article distributed under the terms of the Creative Commons Attribution License (CC BY). The use, distribution or reproduction in other forums is permitted, provided the original author(s) and the copyright owner(s) are credited and that the original publication in this journal is cited, in accordance with accepted academic practice. No use, distribution or reproduction is permitted which does not comply with these terms.
Brain diseases refer to intracranial tissue and organ inflammation, vascular diseases, tumors, degeneration, malformations, genetic diseases, immune diseases, nutritional and metabolic diseases, poisoning, trauma, parasitic diseases, etc. Brain diseases often show disorders of consciousness, sensation, movement, or autonomic nerve dysfunction. There may also be fever, headache, vomiting, and other mental symptoms. Taking Alzheimer’s disease (AD) as an example, the number of patients dramatically increases in developed countries. By 2025, the number of elderly patients with AD aged 65 and over will reach 7.1 million, increasing nearly 29% over the 5.5 million patients of the same age in 2018 (Lynch,
Now, brain diseases are mainly diagnosed by doctors. However, the manual diagnosis requires much time. At the same time, different doctors may have different views on the same examination results, which has brought a lot of trouble to patients.
More and more researchers use computational methods (Wang et al.,
If brain diseases are diagnosed manually, doctors need to spend a lot of time on examination. Sometimes we may encounter the problem that different doctors have different views on the examination results of the same patient. As shown in
Contributions of state-of-the-art methods.
| Method | Contribution |
|---|---|
| Noreen et al. ( |
A multi-level method using two DensNet201 and Inception-v3 was proposed to diagnose early brain tumors. |
| Amin et al. ( |
A model according to the LSTM model method using magnetic resonance images was introduced to classify brain tumors automatically. |
| Amin et al. ( |
A deep learning model was used to predict healthy and unhealthy brain tumor slices. |
| Arunkumar et al. ( |
A new model was introduced to train MRI brain tumors to identify ROI location. |
| Purushottam Gumaste and Bairagi ( |
An algorithm was proposed to extract left and right brain features. This article also introduced different statistical feature extraction methods and used a Support Vector Machine to extract tumor regions from statistical features. |
| Chatterjee and Das ( |
A novel method was proposed for the segmentation of brain images. |
| Bhanothu et al. ( |
A new method based on R-CNN was presented to detect tumors and mark their location. |
| Natekar et al. ( |
Various technologies were compared for brain tumor segmentation models. |
| Aboelenein et al. ( |
The HTTU-Net was proposed for brain tumor cutting. |
| Huang et al. ( |
The DFNN was proposed. The method introduced DFM blocks and combined SE blocks. |
| Hu and Razmjooy ( |
A meta heuristic-based system was presented to detect tumors. |
| Sadad et al. ( |
A novel model according to UNET architecture and ResNet50 as the backbone was proposed for the detection of brain tumors. |
| Kalaiselvi et al. ( |
The PR2G was proposed to detect and segment tumors. |
| Kaplan et al. ( |
Then LBP and αLBP were used to classify the three different types of brain tumors. |
| Khalil et al. ( |
The DA clustering was proposed to improve the accuracy of extracting initial contour points to detect three-dimensional magnetic resonance brain tumors better. |
| Khan et al. ( |
The PART was introduced to detect brain tumors of grade I to grade IV brain tumors. |
| Ma and Zhang ( |
A method was proposed to intelligently detect brain tumors based on a lightweight neural network. |
| Hollon et al. ( |
A new method was proposed for the automatic detection of brain tumors by combining SRH 5–7, CNN, and the label-free optical imaging method. |
| Saba et al. ( |
A new method was proposed to detect brain tumors. The Grasp cut method was used to segment brain tumor symptoms, and VGG-19 was used to obtain features. |
| Sharif et al. ( |
An unsupervised fuzzy set method was introduced for brain tumor segmentation. |
| Xu et al. ( |
A new structure was proposed for the early detection of brain tumors. The new structure was mainly composed of five parts: tumor segmentation, morphology, denoising, feature extraction, and classification. |
| Hemanth et al. ( |
The HSBPN was proposed to segment MR brain tumor images. |
| Nayef et al. ( |
A novel structure was presented for the classification of the MRI dataset. |
| Chen et al. ( |
An improved method was introduced for detecting pathological brains. |
| Shoeibi et al. ( |
A review was presented on the segmentation of the Covid-19 by DL. |
| Shoeibi et al. ( |
A comprehensive survey about the application of DL in the detection of Multiple Sclerosis |
| Sadeghi et al. ( |
A survey was presented on the automatic diagnosis of the SZ by AI. |
| Shoeibi et al. ( |
A comprehensive review was presented on applying the various AI techniques in the diagnosis of Epileptic seizures. |
| Shoeibi et al. ( |
A review of various methods based on DL for automatic diagnosis of SZ by electroencephalogram (EEG) signals was completed. |
| Shoeibi et al. ( |
A new model was proposed to detect Epileptic seizures automatically. The proposed model was based on the DL and the fuzzy theory. |
| Odusami et al. ( |
A method was proposed for the recognition of AD. They tested two CNN models (DenseNet201 and ResNet18) to perform this task. |
| Razzak et al. ( |
A new network (PartialNet) was introduced to detect AD based on MRIs. This network achieved improvements in AD detection. |
| Ashraf et al. ( |
Different CNN models were experimented with to detect AD based on transfer learning. Finally, the fine-tuned DenseNet got the highest accuracy (99.05%). |
To cope with the problems mentioned above, we propose three novel models to classify brain diseases automatically. They are: DenseNet-based Schmidt neural network (DSNN), DenseNet-based random vector functional link (DRVFL), and DenseNet-based extreme learning machine (DELM). We select DenseNet to extract features and use randomized neural networks (RNNs) for classification.
We modify the pre-trained DenseNet. Then, the modified DenseNet is fine-tuned on the dataset. In the DSNN, the last five layers within the fine-tuned DenseNet are substituted by the Schmidt neural network (SNN). In the DRVFL, we select the RVFL (RVFL) to substitute the last five layers of the fine-tuned DenseNet. In the DELM, we choose the extreme learning machine (ELM) to replace the end five layers of the fine-tuned DenseNet.Five-fold cross-validation is used to evaluate the proposed three models: DSNN, DRVFL, and DELM, in terms of aspects (Acc, Sen, Spe, Pre, and F1). We finally get thatDSNN gives the best performance among the three proposed models and overperforms the other six state-of-the-art algorithms. The five main innovations of this study are:
DenseNet is validated as the backbone by experiments showing its superiority to AlexNet, ResNet-18, ResNet-50, and VGG.
DSNN, DRVFL, and DELM are proposed by replacing the last five layers within the fine-tuned DenseNet with three randomized neural networks.
The DSNN gets the best performance among the three proposed models.
The DSNN overperforms the restricted DenseNet and spiking neural network by experiments.
The DSNN is compared with six state-of-the-art algorithms and obtains the best results among the list methods.
The rest of this article is as follows. The dataset is given in Section “Materials". Section “Methodology" discusses the methodology. Section “Results and Discussion" is about the experiment results. We conclude this article in Section “Conclusion".
The dataset is downloaded from the Harvard Medical School website (Johnson and Becker,
Acronym and full explanation.
| Acronym | Full explanation |
|---|---|
| AD | Alzheimer’s disease |
| Acc | Accuracy |
| Avr | Average |
| BN | Batch normalization |
| CNN | Convolution neural network |
| DCNN | Deep convolution neural network |
| DELM | DenseNet-based extreme learning machine |
| DL | Deep learning |
| DRVFL | DenseNet-based random vector functional link |
| DSNN | DenseNet-based Schmidt neural network |
| ELM | Extreme learning machine |
| F1 | F1-score |
| FC | Fully connected |
| ML | Machine learning |
| Pre | Precision |
| RVFL | Random vector functional link |
| RNNs | Randomized neural networks |
| Sen | Sensitivity |
| SNN | Schmidt neural network |
| Spe | Specificity |
| Std | Standard deviation |
The definition of the parameter.
| Parameter | Definition |
|---|---|
| The output of the |
|
| The nonlinear transformation | |
| The given dataset | |
| The input dimension | |
| The output dimension | |
| The weights vector | |
| The bias of the |
|
| The final output weights | |
| The output biases of SNN | |
| The ground-truth label matrix of the dataset | |
| The input matrix | |
| The sigmoid function | |
| The number of hidden nodes | |
| The output matrix of the hidden layer |
The pseudocode of the proposed DSNN is shown in
The pipeline of the proposed DSNN.
Pseudocode of the proposed DSNN.
| Step 1: Load the pre-trained DenseNet. |
| Step 2: Modify the pre-trained DenseNet. |
| Step 2.1 Remove softmax and classification layer from the pre-trainedDenseNet. |
| Step 2.2 Add FC128, ReLU, BN, FC2, softmax, and classification layer. |
| Step 3: Divide the dataset into five groups of the same size and set |
| Step 4: Use the |
| Step 5: Fine-tune the modified DenseNet. |
| Step 5.1: Input is the training set. |
| Step 5.2: Target is the corresponding label. |
| Step 6: Replace the last five layers of the fine-tuned DenseNet with SNN. |
| Step 7: Extract features |
| Step 8: Train the classifier of the DSNN on the extracted features |
| Step 8.1: Input is the extracted features. |
| Step 8.2: The target is the label of the training set. |
| Step 8.3: SNN is the classifier of the DSNN. |
| Step 9: Test the trained DSNN on the test set. |
| Step 10: Report the test classification performance of the trained DSNN. |
| Step 11: Set |
| Step 12: Average test classification performance. |
The CNN models (Albawi et al.,
With the increasing number of network layers in CNN models, researchers are troubled by the problem of gradient vanishing. Batch normalization (BN) alleviates the problem of gradient vanishing to some extent. ResNet (He et al.,
Backbone of the proposed DSNN.
Dense blocks refer to the specific blocks of DenseNet, as shown in
Traditional sequential CNN:
ResNet:
DenseNet:
where [] is the concatenation.
The transition layer is a module that connects different dense blocks. Its primary function is to integrate the features obtained from the previous dense block and reduce its width and height.
Researchers used the ImageNet dataset to pre-train the DenseNet. There are 1,000 output nodes on the pre-trained DenseNet. However, this article only needs two output nodes. We modify the pre-trained DenseNet. The modifications are shown in
After these modifications, we fine-tune the modified DenseNet by the training set. We remove the last five layers of the fine-tuned DenseNet and add SNN to improve the classification performance. In the proposed DSNN, the fine-tuned DenseNet is the feature extraction.
Compared with the pre-trained DenseNet, randomized neural networks (RNNs) have a much shorter training time. In the DSNN, we replace the end five layers of the fine-tuned DenseNet with the RNN: the Schmidt neural network (SNN; Schmidt et al.,
Structure of SNN.
The yellow box is the input, the pink circle represents the hidden nodes, and the green box shows the output. Given
where
The training algorithm of SNN is as follows. The weights vector (
where the sigmoid function is represented as
where the output biases of SNN are
We propose two other models: DRVFL and DELM. The backbone of the two other proposed models is the pre-trained DenseNet. We modify the pre-trained DenseNet in the two proposed models as the “modifications of the pre-trained DenseNet” in the DSNN. We replace the softmax and classification layer of the pre-trained DenseNet with six layers: FC128, ReLU, BN, FC2, softmax, and classification layer. We fine-tune the modified DensNet by the training set. In the DRVFL, we select RVFL (Pao et al.,
Thestructures of
The yellow box represents the input, the pink circle denotes the hidden nodes, and the green box shows the output. The difference between these two RNNs is that there are shortcut connections from the input to the output in RVFL. The calculation steps are similar:
Given
where
For RVFL:
where
where
For ELM:
For RVFL:
where
For ELM:
where
The backbone of these other two proposed models in this article is the same. The difference is that DRVFL chooses RVFL as its classifier, and DELM selects ELM as its classifier.
We define the unhealthy brain as the positive and the healthy brain as the negative. Five indicators are chosen to verify our model: accuracy (Acc), sensitivity (Sen), specificity (Spe), precision (Pre), and F1-score (F1), respectively. Their formulas are shown below:
where the definitions of TP, FN, FP, and TN are the true positive, false negative, false positive, and true negative, respectively.
We modify the hyper-parameter settings of the proposed DSNN. The max-epoch is set to 4 for reducing overfitting problems. We set our mini-batch size to 10 because the dataset is relatively small. According to the experience, the learning rate is 10−4. A hyper-parameter we set in our model is the number of hidden nodes (
The hyper-parameter settings of the proposed DSNN.
| Hyper-parameter | Value |
|---|---|
| Mini-batch size | 10 |
| Max-epoch | 4 |
| Learning rate | 10−4 |
| Number of the hidden nodes V | 400 |
We use five-fold cross-validation to evaluate the proposed DSNN. The classification performance of our model is given in
ROC curve of DSNN.
The classification performance based on five-fold cross-validation (unit: %).
| Methods | Fold | Acc | Sen | Spe | Pre | F1 |
|---|---|---|---|---|---|---|
| DSNN(Ours) | F 1 | 100.00 | 100.00 | 100.00 | 100.00 | 100.00 |
| F 2 | 100.00 | 100.00 | 100.00 | 100.00 | 100.00 | |
| F 3 | 94.87 | 100.00 | 50.00 | 94.59 | 97.22 | |
| F 4 | 100.00 | 100.00 | 100.00 | 100.00 | 100.00 | |
| F 5 | 97.44 | 100.00 | 75.00 | 97.22 | 98.59 | |
| Avr | ||||||
| Std | ±2.05 | ±0.00 | ±20.00 | ±2.17 | ±1.11 | |
| DRVFL(Ours) | F 1 | 100.00 | 100.00 | 100.00 | 100.00 | 100.00 |
| F 2 | 100.00 | 100.00 | 100.00 | 100.00 | 100.00 | |
| F 3 | 89.74 | 100.00 | 0.00 | 89.74 | 94.59 | |
| F 4 | 100.00 | 100.00 | 100.00 | 100.00 | 100.00 | |
| F 5 | 97.44 | 100.00 | 75.00 | 97.22 | 98.59 | |
| Avr | 97.44 | 100.00 | 75.00 | 97.39 | 98.64 | |
| Std | ±3.97 | ±0.00 | ±38.73 | ±3.97 | ±2.10 | |
| DELM(Ours) | F 1 | 100.00 | 100.00 | 100.00 | 100.00 | 100.00 |
| F 2 | 100.00 | 100.00 | 100.00 | 100.00 | 100.00 | |
| F 3 | 92.31 | 100.00 | 25.00 | 92.11 | 95.89 | |
| F 4 | 100.00 | 100.00 | 100.00 | 100.00 | 100.00 | |
| F 5 | 97.44 | 100.00 | 75.00 | 97.22 | 98.59 | |
| Avr | 97.95 | 100.00 | 80.00 | 97.87 | 98.90 | |
| Std | ±2.99 | ±0.00 | ±29.15 | ±3.07 | ±1.60 | |
| Fine-tuned DenseNet | F 1 | 87.50 | 86.11 | 100.00 | 100.00 | 92.54 |
| F 2 | 82.05 | 80.00 | 100.00 | 100.00 | 88.89 | |
| F 3 | 89.74 | 88.57 | 100.00 | 100.00 | 93.94 | |
| F 4 | 85.00 | 83.33 | 100.00 | 100.00 | 90.91 | |
| F 5 | 79.49 | 77.14 | 100.00 | 100.00 | 87.10 | |
| Avr | 84.76 | 83.03 | 100.00 | 100.00 | 90.67 | |
| Std | ±3.67 | ±4.10 | ±0.00 | ±0.00 | ±2.46 | |
| AlexNet-SNN | F 1 | 89.74 | 100.00 | 0.00 | 89.74 | 94.59 |
| F 2 | 89.74 | 100.00 | 0.00 | 89.74 | 94.59 | |
| F 3 | 90.00 | 97.22 | 25.00 | 92.11 | 94.59 | |
| F 4 | 90.00 | 97.22 | 25.00 | 92.11 | 94.59 | |
| F 5 | 89.74 | 97.14 | 25.00 | 91.89 | 94.44 | |
| Avr | 89.84 | 98.32 | 15.00 | 91.12 | 94.56 | |
| Std | ±0.13 | ±1.38 | ±12.25 | ±1.13 | ±0.06 | |
| ResNet-18-SNN | F 1 | 100.00 | 100.00 | 100.00 | 100.00 | 100.00 |
| F 2 | 97.50 | 100.00 | 75.00 | 97.30 | 98.63 | |
| F 3 | 100.00 | 100.00 | 100.00 | 100.00 | 100.00 | |
| F 4 | 94.87 | 100.00 | 50.00 | 94.59 | 97.22 | |
| F 5 | 94.87 | 97.14 | 75.00 | 97.14 | 97.14 | |
| Avr | 97.45 | 99.43 | 80.00 | 97.81 | 98.60 | |
| Std | ±2.29 | ±1.14 | ±18.71 | ±2.03 | ±1.26 | |
| ResNet-50-SNN | F 1 | 95.00 | 94.44 | 100.00 | 100.00 | 97.14 |
| F 2 | 100.00 | 100.00 | 100.00 | 100.00 | 100.00 | |
| F 3 | 97.44 | 97.14 | 100.00 | 100.00 | 98.55 | |
| F 4 | 95.00 | 100.00 | 50.00 | 94.74 | 97.30 | |
| F 5 | 100.00 | 100.00 | 100.00 | 100.00 | 100.00 | |
| Avr | 97.49 | 98.32 | 90.00 | 98.95 | 98.60 | |
| Std | ±2.24 | ±2.23 | ±20.00 | ±2.10 | ±1.24 | |
| VGG-SNN | F 1 | 97.50 | 100.00 | 75.00 | 97.30 | 98.63 |
| F 2 | 87.50 | 94.44 | 25.00 | 91.89 | 93.15 | |
| F 3 | 94.87 | 97.14 | 75.00 | 97.14 | 97.14 | |
| F 4 | 89.74 | 100.00 | 0.00 | 89.74 | 94.59 | |
| F 5 | 87.18 | 88.57 | 75.00 | 96.88 | 92.54 | |
| Avr | 91.36 | 96.03 | 50.00 | 94.59 | 95.21 | |
| Std | ±4.12 | ±4.26 | ±31.62 | ±3.16 | ±2.33 | |
| Restricted DenseNet-SNN | F 1 | 94.87 | 94.29 | 100.00 | 100.00 | 97.06 |
| F 2 | 100.00 | 100.00 | 100.00 | 100.00 | 100.00 | |
| F 3 | 97.37 | 97.06 | 100.00 | 100.00 | 98.51 | |
| F 4 | 94.87 | 100.00 | 50.00 | 94.59 | 97.22 | |
| F 5 | 100.00 | 100.00 | 100.00 | 100.00 | 100.00 | |
| Avr | 97.42 | 98.27 | 90.00 | 98.92 | 98.56 | |
| Std | ±2.09 | ±1.83 | ±1.97 | ±2.09 | ±1.69 |
The bold values are results of our proposed model.
The classification performances of DRVFL and DELM based on the five-fold cross-validation are shown in
Modelcomparison.
We compare the proposed DSNN with the fine-tuned DenseNet. The classification performance of the fine-tuned DenseNet is given in
DenseNet has too many layers and parameters and is prone to meet overfitting problems because our dataset is relatively small. The structure of SNN is simple and has only three layers. What’s more, there are fewer parameters in SNN, which is not easy to produce overfitting problems. So, our method achieves better accuracy than fine-tuned DenseNet.
We test the performance of the proposed DSNN with different backbones. These backbones are AlexNet, ResNet-18, ResNet-50, and VGG, respectively. The classification performances of the proposed DSNN with different backbones are shown in
DenseNet as the backbone model achieves the best results compared with other backbones. The reason is that DenseNet can reduce gradient vanishing problems better than other CNN models by establishing dense connectivity between all front and rear layers. There are too many parameters in VGG and AlexNet. There are 138M parameters for VGG and 61M parameters for AlexNet. However, there are only 20M parameters in DenseNet. More epoch is needed to converge for VGG and AlexNet. Nevertheless, to prevent overfitting problems, we set the max-epoch to 4. Therefore, DenseNet obtains better performance than VGG and AlexNet. Compared with ResNet, dense connections in the layers can provide more supervision information so that DenseNet can produce better classification performance. DenseNet has also shown its superiority in image learning in other studies, such as Ker et al. (
We limit the number of connections in the DenseNet block. Each layer is only connected to the previous layer in the last block. The results are shown in
We compare the proposed DSNN with the spiking neural network (Yaqoob and Wróbel,
The final result of the spiking neural network (unit: %).
| Model | Acc | Sen | Spe | Pre | F1 |
|---|---|---|---|---|---|
| Spiking neural network | 86.05 | 100.00 | 0.00 | 96.05 | 92.50 |
The bold values are results of our proposed model.
It is significant to explain the DCNNs because it is difficult for researchers to figure out how DCNNs make predictions. We can visualize the attention of DCNNs by the Gradient-weighted class activation mapping (Grad-CAM). We present the raw images and heatmap images in
Explainability of the proposed DSNN.
The blue region is the lowest attention in Grad-CAM. Based on the Grad-CAM, we can conclude that DSNN can classify brain diseases in MRI. Also, some other studies have proven that the Grad-CAM efficiently visualizes the attention of DCNNs, such as Chattopadhay et al. (
We compare the proposed DSNN with other state-of-the-art methods. These state-of-the-art methods are: ANN (Arunkumar et al.,
Comparison with other state-of-the-art methods (unit: %).
Comparison with other state-of-the-art methods (unit: %).
| Methods | Sen | Spe | Pre | Acc | F1 |
|---|---|---|---|---|---|
| ANN (Arunkumar et al., |
89.00 | - | - | 92.14 | - |
| PR2G (Kalaiselvi et al., |
98.46 | - | - | 83.90 | - |
| SRH + CNNs (Hollon et al., |
- | - | - | - | |
| BPNN (Hemanth et al., |
57.54 | 54.50 | 91.71 | 57.23 | 70.72 |
| LVQNN (Nayef et al., |
59.94 | 61.00 | 93.08 | 60.05 | 72.92 |
| LRC (Chen et al., |
100.00 | 58.50 | 95.47 | 95.74 | 97.68 |
| DSNN (Ours) |
Our model is an effective method to classify brain diseases based on the comparison results. The proposed DSNN can achieve these good results because deep learning is used to extract features, and SNN is used for classification.
Three novel models are proposed to automatically classify brain diseases in this article. The proposed models are DSNN, DRVFL, and DELM. The DSNN gets the best performance among the three proposed models in terms of classification performance. The backbone of the proposed DSNN is the pre-trained DenseNet. We modify the pre-trained DenseNet. Then, the modified DenseNet is fine-tuned on the dataset. The last five layers within the fine-tuned DenseNet are substituted by the Schmidt neural network (SNN). In the proposed DSNN, the fine-tuned DenseNet plays the role of feature extraction. The extracted features train the SNN. We evaluate the proposed DSNN by using five-fold cross-validation. The accuracy, sensitivity, specificity, precision, and F1-score of the proposed DSNN on the test set are 98.46% ± 2.05%, 100.00% ± 0.00%, 85.00% ± 20.00%, 98.36% ± 2.17%, and 99.16% ± 1.11%, respectively. The proposed DSNN is compared with other state-of-the-art methods and obtains the best results among the list methods. Our model obtaining the best performance can conclude that DSNN is an effective model for classifying brain diseases.
Although the proposed model gets good results, this article still has some shortcomings. (1) The dataset is relatively small. (2) We divide the datasets into two categories. However, there are many kinds of brain diseases.
We will collect more data to test the proposed model in the future. Then, we will try to classify multiple brain diseases. What’s more, we will do more research on brain segmentation. We will try more new deep learning methods, such as VIT, attention learning, etc.
Publicly available datasets were analyzed in this study. This data can be found here: https://www.med.harvard.edu/aanlib.
ZZ: conceptualization, software, data curation, writing—original draft, writing—review and editing, visualization. SL: conceptualization, software, data curation, writing—review and editing. S-HW: methodology, software, validation, investigation, resources, writing—review and editing, supervision, and funding acquisition. JG: methodology, validation, formal analysis, resources, writing—original draft, writing—review and editing, supervision. Y-DZ: methodology, formal analysis, investigation, data curation, writing—original draft, writing—review and editing, visualization, supervision, project administration, and funding acquisition.
The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.
All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.
The article was partially supported by Hope Foundation for Cancer Research, UK (RM60G0680); Royal Society International Exchanges Cost Share Award, UK (RP202G0230); Medical Research Council Confidence in Concept Award, UK (MC_PC_17171); British Heart Foundation Accelerator Award, UK (AA/18/3/34220); Sino-UK Industrial Fund, UK (RP202G0289); Global Challenges Research Fund (GCRF), UK (P202PF11); LIAS Pioneering Partnerships award, UK (P202ED10); Data Science Enhancement Fund, UK (P202RE237); Guangxi Key Laboratory of Trusted Software (kx201901).