Pressure wounds are skin injuries that occur due to prolonged inactivity of the patient. They reduce blood flow. They cause muscle and skin ischemia, which leads to necrosis and cell death. The automatic classification of pressure sores according to international principles allows rapid recommendation of appropriate treatment methods. In this way, life quality of patients will be increased by ensuring that the treatments to be applied are efficient. Diabetes foot wounds are among the current serious and chronic diseases and are among the wound types that threaten life and reduce the quality of life. In the literature, there are mostly studies that make binary classification (ulcer-nonulcer) showing which type pressure and diabetic foot ulcers belong to. In this study, the impact of recently popular deep learning methods in image processing for tissue classification of pressure and diabetic wound images is discussed. Deep Convolutional Neural Network (CNN) architectures were used to determine which of the granule, necrotic and slough classes wound images belong to, according to their characteristics on the tissue. In this study, in which 19 CNN architectures were discussed, CNN architectures that give successful results on the dataset were observed. In addition, the effect of Maximum Epoch Number (ME), Initial learning rate (LR) and Minibatch size (MBS) parameter values used in the architectures were also researched. In the experiments performed on 1045 pressure and 1045 diabetic wound images collected from different patients from the Wound Care Unit of Karaman Education and Research Hospital, the best results were obtained with the AlexNet architecture with an Accuracy value of 95.83%. Optimum values of ME, LR and MBS parameter values were also found in parameter optimization tests. Comparisons were made with the results of studies on similar wound datasets in the literature, and it was shown that AlexNet and VGG16 architectures achieved competitive results.
Research Article
Classification of Pressure and Diabetic Chronic Wound Tissue Images with Deep Learning Methods
https://doi.org/10.21203/rs.3.rs-2082678/v1
This work is licensed under a CC BY 4.0 License
Version 1
posted
You are reading this latest preprint version
Wound Image Classification
Convolutional Neural Networks
Deep Learning
Pressure Ulcers
Diabetic Foot Ulcers
A pressure wound is local injury to the skin and/or underlying soft tissue, usually with a bony prominence or in association with a medical device or other device. Pressure injuries can occur as intact skin or open injuries and can be painful. Such type of injuries are the result of intense and prolonged pressure (NPUAP- National Pressure Ulcer Advisory Panel, 2014) [1], [2], [3].
Pressure injuries reduce the quality of life due to pain and infection, prolong hospital stay and increase the cost of care. The time devoted to preventing pressure injuries is shorter and less costly than that for their care and treatment. Often it is resistant to recovery and requires long-term medical care [4].
In the same way, the evaluation of diabetic foot wounds is a topical issue in research in terms of health economics and trends in accelerating treatment [5]. It is estimated that 536.6 million people worldwide have diabetes in 2021, and it will increase rapidly to reach 783.2 million people in 2045. Clinically, it is important to enabling earlier diagnosis and earlier initiation of treatment during the asymptomatic stage [6, 7]. Automatic classification and tracking of wound tissues with high accuracy have great importance for the correct definition of wound care to be applied [8]. Chronic wound tissue can be classified as red-colored granular tissue, yellow-colored slough tissue, and black-colored necrotic tissue as in Figure 1.
In recent years, studies on image classification by using deep learning methods have been increasing day by day [9]. Studies can generally be grouped as modifications made on CNN models, investigating the effect on model parameters, and determining the one with the highest effect by using models together (ensemble). In addition, in some studies in the literature, there are also studies on the dataset that is collected exclusively for the study instead of a certain image dataset.
[10] performed Binary classification on 1335 wound images collected from patients in Palo Alto VA Hospital and the Washington University Medical Center in St. Louis. With the method they call DeepWound, a positive or negative output of a wound image is produced for each of the 9 classes they have determined. [11], on the other hand, studied 400 images belonging to 4 classes, 100 of each of the venous, diabetic, pressure, and surgical wound types, which they collected at the AZH Wound and Vascular Center in Milwaukee, Wisconsin. Ensemble Deep performed binary Classification with Surgical and Venous classes, and 3-class Classification with Surgical, venous and Diabetic classes with a CNN-based classifier. In the tests with pictures of 4 classes, the success rate is quite low with 68.69%.
[12] proposed a CNN architecture, which they named DFU_QUTNet, on 754 foot pictures they collected from patients in Iraq Nasiriyah Hospital Diabetic Center. In the study, the success of the images on the normal or abnormal classes was measured by binary classification of the wounds. [13] studied a total of 397 images from Lancashire Teaching Hospitals, of which 292 were from diabetic foot wounds (abnormal) and 105 were from healthy patients. The binary classification was performed on the wounds and they obtained 92.5% accuracy for normal and abnormal classes. [14] determined whether diabetic foot wounds were infection or ischemia on 1459 new datasets they collected from Lancashire Teaching Hospitals. They used InceptionV3, ResNet50 and InceptionResNetV2 architectures. They achieved a success of 90.3% in binary classification of whether there is ischemia in the images and 72.7% in binary classification ofwhether there is infection. [15], studied a total of 397 images from Lancashire Teaching Hospitals, of which 292 were from diabetic foot wounds (abnormal) and 105 were from healthy patients. They compared the results obtained from the proposed CNN architecture named DFU SPNet with the studies of [12] and [13].
When the literature is examined, it is seen that there are mostly studies on binary classification of existence or the lack thereof. Furthermore, there are studies that indicate which wound type the wound images belong to, rather than the classification of tissue type. It is also observed that CNN architectures are being used in some studies. Most studies generate their own datasets themselves.
In this study, experimental analyzes were carried out to ensure that the deep learning approaches used can perform tissue analysis in different wound types with high performance on a newly produced dataset for this study. Initially, it is aimed to measure the success of 19 CNN architectures with constant initialized parameters on the existing dataset. In the next step, the effects of parameter values used in CNN architectures on the results are discussed. This study aims to find the epoch number, batch size, and learning rate parameters that reach the best success in the classification of wound images. The dataset, preprocessing steps, and information about the architectures and parameters used in this study are explained in Section 2. Information about the dataset used is also shared in this section. Details of the experiments performed are presented in Section 3. The discussion of the achievements of the architectures and the test results is given in Section 4. Finally, the paper was concluded with the results section.
2.1 Dataset, Preprocessing and Augmentation
In this study, 2090 images collected in the categories of granule, necrotic, and slough from different patients in Karaman Education and Research Hospital in Turkey were used. Pictures were obtained with iPhone 7 Plus and 48mp Android operating system mobile phone and Ipad Pro tablet from a maximum distance of 30 cm to cover the entire wound area. The pictures are generally at 1200x1600 resolution and 1045 of them are pressure wounds and 1045 of them are diabetic foot wounds. Examples of the pictures in the dataset are given in Figure 2.
In the pre-processing step, the images were resized as 224x224, 227x227, 256x256, 299x299, and 331x331 pixels according to the image input size of each CNN model. In addition, in order to avoid overfitting during the training phase, data augmentation was done with rotation, reflection, brightness, and Gaussian noise methods. The condition of the samples in Figure 2 after augmentation is given in Figure 3.
The dataset information used in the paper is presented in Table 1.
Table 1Distribution of the original data collected and post-enrichment
|
Original |
After data augmentation |
|
|
Pressure Wound |
1045 (350 granular, 350 necrotic and 345 slough) |
4180 (1400 granular, 1400 necrotic and 1380 slough) |
|
Diabetic Foot Wound |
1045 (350 granular, 350 necrotic and 345 slough) |
4180 (1400 granular, 1400 necrotic and 1380 slough) |
2.2 The CNN Architectures Used
Deep learning is a sub-title of machine learning and it is used in computer vision tasks. Recently, deep learning architectures have come to the fore with their achievements in image analysis. Different deep learning architectures can show successful results in the datasets consisting of different images studied in the literature. In addition, parameter selection is also important in utilizing deep learning architectures. The classifier deep learning architectures used in this paper are introduced below.
Googlenet: This architecture is a CNN architecture with a depth of 22 layers. It was recommended by Google researchers for the ILSVRC 2014 competition. It has an image input size of 224x224. In this architecture used for classification, the output value is adjusted according to the number of classes in a certain dataset. It came first in the ILSVRC2014 competition with an error rate of 6.7%. [16].
Densenet201: This CNN architecture proposed by Huang et al.has a depth of 201 layers [17]. This network with 224x224 image input size has been developed with inspiration from Resnet architecture. To further improve the flow of information between layers with Dense connectivity, the connections of each layer establish connections to the layers that follow it [17].
Mobilenetv2: It is a CNN architecture recommended for mobile devices [18]. This architecture recommended by Google has an image input size of 224x224. This architecture, which has a depth of 53 layers, is the advanced version of the mobilenetv1 architecture.
Resnet18: It is the model recommended by Microsoft and achieved 96.4% accuracy in the ImageNet 2016 competition [19]. Resnet18 architecture, which has an image input size of 224x224, has a depth of 18 layers.
Resnet50: Thisarchitecture, which is proposed by He et al. is the winner of the ILSVRC 2015 competition and has an image input size of 224x224 [19] It has a depth of 50 residual networks. It consists of an Identity Block structure consisting of Convolution layers. It is a deep residual network with an identity connection between layers. In this way, it is aimed to eliminate the vanishing gradient problem.
Resnet101: This architecture, proposed by He et al. has an image input size of 224x224. It has a depth of 101 layers [19].
Shufflenet: This architecture, proposed by Zhang et al., is for mobile devices with limited-limit computing power, and it has an image input size of 224x224 with a depth of 50 layers [20]. It includes a new shuffle operation that is used to assist the information flow between features channels.
Nasnetmobile: Neural Architecture Search Network (Nasnet) developed by the Google team has an image input size of 224x224. It is an architecture built from basic building blocks that are optimized using Reinforcement learning. This architecture consists of 12 blocks that have 5.3 million parameters [21].
Efficientnetb0: The input resolution is 224x224 and has a depth of 82 layers. The architecture recommended by Tan and Le is an effective architecture in ImageNet and image classification transfer learning tasks [22].
Vgg16: It is the model recommended by Oxford and achieved 92.7% accuracy in the ImageNet 2013 competition [23]. It has a depth of 16 layers, and an image input size of 224x224. High kernel sizes are reduced compared to Alexnet. The VGGNet architecture uses 3x3 filters in all its layers and the Convolution-ReLU layers before the pooling layer. In the VGG architecture, as in other deep architectures, while the height and width dimensions of the matrices decrease from the input layer to the output, the depth value increases.
Vgg19: It is a CNN architecture with a depth of 19 layers. The model uses an image input size of 224×224. The only difference from the 16-layer Vgg16 architecture is the number of layers [23].
Squeezenet: It is an architecture with 50 layers of depth and 224x224 input resolution proposed by Iandola et al. [24]. It uses 1x1 filters that reduce network weights in order to increase the performance of traditional CNN architectures. Thanks to the distributed layers, it works faster by reducing the computational cost compared to Alexnet architecture.
Alexnet: It is a convolutional neural network architecture developed by Krizhevsky et al. [25] with an input image size of 227x227. AlexNet with 8 layers of depth is the winner of the ILSVRC 2012 competition. The first 5 layers consist of convolutional layers and the last 3 layers are fully connected layers. There are also Pooling and Activation layers between the layers. It has a prominent place among CNN models [26].
Darknet19: This architecture is a CNN architecture with a depth of 19 layers. It has an image input size of 256x256. It is used as the backbone model of the YOLOv2 model, which is used in object identification. In the architecture that starts with a 3x3 filter, the number of channels is doubled after each pool layer [27].
Darknet53: This architecture is a CNN architecture with a depth of 53 layers. It has an image input size of 256x256 [28]. The advanced model of the previous version Darknet19 is used as the backbone model of the YOLOv3 object completion model.
Inceptionv3: Googlenet (Inceptionv1), which was created as an initial model, is built on CNN architecture [16]. Szegedy et al. developed Inceptionv3 as a new version of the initial model [29]. Inception_v3 has more parameters than inception_V1 and inception_V2. The same input tensor is used with multiple filters, and the outputs are combined and sent to the next inception module. It has a parallel convolutional layer block with 3 different sizes of filters (1x1, 3x3, 5x5). In addition to this, 3x3 max pooling is done. The architecture with 48 layers has an image input size of 299x299.
Xception : This CNN architecture proposed by Chollet consists of 71 layers and has an image input size of 299x299 [30]. This architecture with in-depth convolutions is an enhanced version of the inception architecture.
Inceptionresnetv2: It is a CNN architecture proposed by Szegedy et al. and trained with more than one million images from the ImageNet database [31]. The model uses an image input size of 299×299, but outputs a list of class possibilities. It is a hybrid technique where Inception and Resnet are combined.
Nasnetlarge: It is a convolutional architecture trained on over a million images from the Imagenet database. It has an image input size of 331x331 [21].
2.3 Evaluation Metrics for Classification
In order to measure the performance of the classification models, the actual image value on a pixel basis is compared with the result values predicted by the model. Used in evaluation metric formulas, TP, TN, FP, and FN indicates true positive (correctly validated sample), true negative (correctly rejected sample), false positive (falsely validated sample), and false negative (falsely rejected sample), respectively. Detailed information about the specified evaluation metrics can be obtained from the study [32]. The evaluation metrics used in this study are given in Table 2.
Accuracy is the ratio of the number of correctly classified samples to the total number of samples. It is the overall performance of the model.
Sensitivity, (TPR=True Positive Rate) refers to the true positive rate and summarizes how well the positive class was predicted. It is the metric that shows how many of the samples that should be predicted positively are predicted positively.
Specificity, (TNR= True Negative Rate) refers to the true negative rate and is the rate that indicates how well the negative class is predicted.
Precision, (PPV= Positive Predictive Value) is a metric that shows how many of the samples classified correctly as a result of the classification are actually positive.
The F1 Score is the harmonic mean of the Sensitivity and Precision metrics. If any of these two metrics have a low value, it prioritizes that value, providing a better model evaluation.
All evaluation metrics are interpreted as a percentage value.
In this paper, the classification successes of 19 deep learning architectures on the dataset of diabetic and pressure wound images collected from patients through a specialist are discussed. The measurement metrics in
Table 2 were used to compare the model classification successes.
A computer with the technical specifications of Xeon Silver 4114 2.2GHz processor, 32 GB RAM, NVIDIA Quadro P5000 (x2) GPU was used to get experimental results, in the training and testing processes. Python language, and Keras, Tensorflow, Pytorch, Tensorboard libraries were used in all the experiments.
3.1. Parameter Optimization
In order to obtain a successful classification of wound images, different deep learning architectures were tested with different parameters, and the architectures' conditions were determined. The effects of the maximum epoch, minibatch size, and learning rate, which are the CNN model training parameters, on wound classification have been demonstrated with experimental results. In experimental studies, 5 measurement metrics, namely Accuracy, Dice (f1 score), sensitivity (recall), specificity, and precision, were used to measure and compare the performances of the models. Information about the pre-trained networks used in experimental studies is given in Table 3.
Table 3Used Architectures
|
id |
CNN Architectures |
Depth |
Parameters (Millions) |
Image Input Size |
|
1 |
googlenet |
22 |
7,00 |
224-by-224 |
|
2 |
densenet201 |
201 |
20,00 |
224-by-224 |
|
3 |
mobilenetv2 |
53 |
3,50 |
224-by-224 |
|
4 |
resnet18 |
18 |
11,70 |
224-by-224 |
|
5 |
resnet50 |
50 |
25,60 |
224-by-224 |
|
6 |
resnet101 |
101 |
44,60 |
224-by-224 |
|
7 |
shufflenet |
50 |
1,40 |
224-by-224 |
|
8 |
nasnetmobile |
12 |
5,30 |
224-by-224 |
|
9 |
efficientnetb0 |
82 |
5,30 |
224-by-224 |
|
10 |
vgg16 |
16 |
138,00 |
224-by-224 |
|
11 |
vgg19 |
19 |
144,00 |
224-by-224 |
|
12 |
squeezenet |
18 |
1,24 |
227-by-227 |
|
13 |
alexnet |
8 |
61,00 |
227-by-227 |
|
14 |
darknet19 |
19 |
20,80 |
256-by-256 |
|
15 |
darknet53 |
53 |
41,60 |
256-by-256 |
|
16 |
inceptionv3 |
48 |
23,90 |
299-by-299 |
|
17 |
xception |
71 |
22,90 |
299-by-299 |
|
18 |
inceptionresnetv2 |
164 |
55,90 |
299-by-299 |
|
19 |
nasnetlarge |
* |
88,90 |
331-by-331 |
In deep learning architectures, the values of the parameters directly affect the results. While fixed values are used in some studies, appropriate values are determined over a sample model in others. Therefore, the initial aim was to find the epoch number, batch size, and learning rate that best achieves success in the classification of wound images. In the parameter determination experiments, the decay (drop factor, 0.5) and drop period (5) parameters were kept constant in the dataset classification, which was separated by the hold-out method as 70% train 30% test data. Epoch numbers, batch size, and learning rate parameters were initially randomly determined. Primarily, experiments were conducted using 10 epochs, 32 batch size, and 10e-4 learning rate parameters. The parameter values in Table 4 were applied to all models given in Table 3 in the same way. The success of the models in the evaluation metrics is given in Table 5.
Table 4Used parameter values
|
Maximum epoch (ME) |
Constant Initialized |
|
Initial learning rate (LR) |
Constant Initialized |
|
Drop factor (DF) |
0.5 |
|
Drop period (DP) |
5 |
|
Minibatch size (MBS) |
Constant Initialized |
|
Optimizer |
Stochastic Gradient Descent with Momentum (SGDM) |
Table 5Performance of Models with Constant Parameters
|
id |
CNN Architectures |
Accuracy |
Sensitivity |
Specificity |
Precision |
F-Score |
|
1 |
googlenet |
75,0000 |
76,4706 |
97,0588 |
75,0000 |
64,7059 |
|
2 |
densenet201 |
72,9167 |
71,4706 |
97,0588 |
66,6667 |
61,9048 |
|
3 |
mobilenetv2 |
72,9167 |
71,4706 |
80,6452 |
59,0909 |
66,6667 |
|
4 |
resnet18 |
81,2500 |
71,4706 |
88,2353 |
69,2308 |
66,6667 |
|
5 |
resnet50 |
75,0000 |
64,7059 |
94,1176 |
66,6667 |
58,8235 |
|
6 |
resnet101 |
79,1667 |
70,5882 |
88,2353 |
69,2308 |
61,5385 |
|
7 |
shufflenet |
68,7500 |
47,0588 |
88,2353 |
42,8571 |
40,0000 |
|
8 |
nasnetmobile |
68,7500 |
52,9412 |
85,2941 |
44,4444 |
41,8605 |
|
9 |
efficientnetb0 |
68,7500 |
76,4706 |
85,2941 |
54,1667 |
63,4146 |
|
10 |
vgg16 |
77,0833 |
70,5882 |
88,2353 |
63,6364 |
60,0000 |
|
11 |
vgg19 |
81,2500 |
82,3529 |
90,3226 |
70,0000 |
70,0000 |
|
12 |
squeezenet |
66,6667 |
70,5882 |
82,3529 |
52,1739 |
60,0000 |
|
13 |
alexnet |
89,5833 |
71,4286 |
97,0588 |
90,9091 |
80,0000 |
|
14 |
darknet19 |
83,3333 |
70,5882 |
94,1176 |
70,0000 |
66,6667 |
|
15 |
darknet53 |
64,5833 |
64,2857 |
80,6452 |
45,4545 |
51,4286 |
|
16 |
inceptionv3 |
62,5000 |
41,5966 |
70,9677 |
62,9032 |
37,2727 |
|
17 |
xception |
70,8333 |
64,7059 |
97,0588 |
50,0000 |
46,8085 |
|
18 |
inceptionresnetv2 |
58,3333 |
35,2941 |
79,4118 |
30,0000 |
32,4324 |
|
19 |
nasnetlarge |
72,9167 |
64,7059 |
85,2941 |
54,5455 |
56,4103 |
When the evaluation metrics in Table 5 are examined, the best results for Accuracy, Specificity, Precision, and F-Score were obtained in the AlexNet architecture. For the Sensitivity, Specificity, and Precision metrics, the GoogleNet architecture achieved the second-best results. For the specificity metric, densenet201 was equally achieved with xception architectures, AlexNet and GoogleNet, but measured less in other metrics. Therefore, AlexNet and GoogleNet architectures were chosen for parameter optimization, which produced better values among all models.
In order to find the best other parameters, tests were conducted with different epoch numbers and minibatch size values in the 2nd test scenario. For this reason, these two architectures were run with 10,25,50,100,150,200,250, and 300 epochs, and the obtained results are shown in Table 6.
Table 6Results of GoogleNet and AlexNet Architectures in Different Epochs
|
Epoch Number |
Accuracy |
Sensitivity |
Specificity |
Precision |
F-Score |
|
|
GoogleNet |
10 |
83,3333 |
88,2353 |
93,1176 |
80,0000 |
75,0000 |
|
25 |
77,0833 |
57,1429 |
90,3226 |
75,0000 |
62,0690 |
|
|
50 |
79,1667 |
71,4286 |
83,8710 |
64,2857 |
66,6667 |
|
|
100 |
77,0833 |
76,4706 |
85,2941 |
61,5385 |
66,6667 |
|
|
150 |
85,4167 |
82,3529 |
94,1176 |
81,8182 |
75,6757 |
|
|
200 |
85,4167 |
64,7059 |
87,0588 |
88,8889 |
69,5652 |
|
|
250 |
81,2500 |
70,5882 |
87,0968 |
75,0000 |
72,7273 |
|
|
300 |
72,9233 |
76,4629 |
83,8710 |
72,2200 |
74,2857 |
|
|
Epoch Number |
Accuracy |
Sensitivity |
Specificity |
Precision |
F-Score |
|
|
AlexNet |
10 |
89,5833 |
71,4286 |
97,0588 |
90,9091 |
80,0000 |
|
25 |
85,4167 |
78,5714 |
93,5484 |
84,6154 |
75,8621 |
|
|
50 |
83,3333 |
85,7143 |
90,3226 |
76,9231 |
75,0000 |
|
|
100 |
87,5000 |
76,4706 |
97,0588 |
90,0000 |
75,0000 |
|
|
150 |
95,8333 |
94,1176 |
97,0588 |
94,1176 |
94,1176 |
|
|
200 |
81,2500 |
70,5882 |
91,1765 |
72,7273 |
66,6667 |
|
|
250 |
83,3333 |
76,4706 |
88,2353 |
76,4706 |
76,4706 |
|
|
300 |
87,5000 |
88,2353 |
97,0588 |
90,0000 |
76,9231 |
When the evaluation metrics inTable 6 are examined, GoogleNet architecture achieved the best results in 150 epochs with 3 metrics: Accuracy, Specificity, and F-Score, while AlexNet architecture achieved the best results in 150 epochs in all metrics. In the AlexNet architecture, 10,100,150, and 300 in the Specificity metric are equal in epochs. In these two models, 150 epoch numbers, which produced the highest number of values in the best results, were used in the other tests.
In the next step, 150 epochs (ME) of both models were run in order to find effective minibatch sizes (MBS) and learning rates (LR) in obtaining the best results in both models. Both architectures were tested with 3 different LR parameters (10e-3 ,10e-4 , 10e-5) by taking 2,4,8,16,32,64,128 and 256 MBS. The results obtained for the AlexNet architecture are given in Table 7, and the results obtained for the GoogleNet Architecture are given in Table 8.
Table 7Results of the AlexNet Architecture in Different MBS and LR parameters (ME=150)
|
LR |
MBS |
Accuracy |
Sensitivity |
Specificity |
Precision |
F-Score |
|
10e-3 |
2 |
56,9444 |
33,3333 |
66,6667 |
35,4167 |
52,3077 |
|
4 |
56,9444 |
33,3333 |
66,6667 |
35,4167 |
52,3077 |
|
|
8 |
56,9444 |
33,3333 |
66,6667 |
35,4167 |
52,3077 |
|
|
16 |
56,9444 |
33,3333 |
66,6667 |
35,4167 |
52,3077 |
|
|
32 |
68,7500 |
52,9412 |
76,4706 |
47,3684 |
50,0000 |
|
|
64 |
93,7500 |
78,5714 |
100,0000 |
100,0000 |
88,0000 |
|
|
128 |
83,3333 |
71,4286 |
88,2353 |
71,4286 |
71,4286 |
|
|
256 |
83,3333 |
92,8571 |
87,0968 |
65,0000 |
76,4706 |
|
|
10e-4 |
2 |
70,8333 |
94,1176 |
66,6667 |
50,0000 |
50,7937 |
|
4 |
77,0833 |
64,7059 |
83,8710 |
68,7500 |
66,6667 |
|
|
8 |
85,4167 |
85,7143 |
90,3226 |
78,5714 |
77,4194 |
|
|
16 |
91,6667 |
88,2353 |
97,0588 |
91,6667 |
84,6154 |
|
|
32 |
95,8333 |
94,1176 |
97,0588 |
94,1176 |
94,1176 |
|
|
64 |
79,1667 |
71,4286 |
87,0968 |
66,6667 |
66,6667 |
|
|
128 |
91,6667 |
94,1176 |
96,7742 |
92,3077 |
88,8889 |
|
|
256 |
89,5833 |
82,3529 |
94,1176 |
84,6154 |
81,4815 |
|
|
10e-5 |
2 |
81,2500 |
70,5882 |
88,2353 |
69,2308 |
66,6667 |
|
4 |
83,3333 |
82,3529 |
96,7742 |
90,9091 |
71,7949 |
|
|
8 |
91,6667 |
78,5714 |
97,0588 |
91,6667 |
84,6154 |
|
|
16 |
91,6667 |
88,2353 |
100,0000 |
100,0000 |
83,3333 |
|
|
32 |
83,3333 |
64,7059 |
91,1765 |
75,0000 |
69,2308 |
|
|
64 |
77,0833 |
92,8571 |
87,0968 |
66,6667 |
70,2703 |
|
|
128 |
87,5000 |
76,4706 |
100,0000 |
100,0000 |
72,7273 |
|
|
256 |
72,9167 |
70,5882 |
90,3226 |
62,5000 |
52,1739 |
Table 8Results of the GoogleNet Architecture in Different MBS and LR parameters (ME=150)
|
LR |
MBS |
Accuracy |
Sensitivity |
Specificity |
Precision |
F-Score |
|
10e-3 |
2 |
83,3333 |
82,3529 |
83,8710 |
73,6842 |
77,7778 |
|
4 |
89,5833 |
94,1176 |
96,7742 |
88,8889 |
83,8710 |
|
|
8 |
91,6667 |
92,8571 |
93,5484 |
85,7143 |
86,6667 |
|
|
16 |
83,3333 |
76,4706 |
91,1765 |
75,0000 |
70,2703 |
|
|
32 |
89,5833 |
88,2353 |
94,1176 |
87,5000 |
84,8485 |
|
|
64 |
85,4167 |
76,4706 |
90,3226 |
81,2500 |
78,7879 |
|
|
128 |
81,2500 |
88,2353 |
97,0588 |
83,3333 |
70,9677 |
|
|
256 |
85,4167 |
76,4706 |
82,7957 |
74,0476 |
70,2703 |
|
|
10e-4 |
2 |
85,4167 |
71,4286 |
91,1765 |
76,9231 |
74,0741 |
|
4 |
81,2500 |
70,5882 |
78,4946 |
71,0648 |
66,6667 |
|
|
8 |
81,2500 |
82,3529 |
88,2353 |
70,0000 |
75,6757 |
|
|
16 |
79,1667 |
76,4706 |
88,2353 |
68,4211 |
72,2222 |
|
|
32 |
85,4167 |
82,3529 |
94,1176 |
81,8182 |
75,6757 |
|
|
64 |
83,3333 |
82,3529 |
94,1176 |
80,0000 |
77,7778 |
|
|
128 |
75,0000 |
64,7059 |
94,1176 |
66,6667 |
55,5556 |
|
|
256 |
85,4167 |
71,4286 |
91,1765 |
76,9231 |
74,0741 |
|
|
10e-5 |
2 |
70,8333 |
64,7059 |
87,0968 |
63,6364 |
50,0000 |
|
4 |
85,4167 |
82,3529 |
94,1176 |
81,8182 |
75,6757 |
|
|
8 |
77,0833 |
64,7059 |
85,2941 |
64,7059 |
64,7059 |
|
|
16 |
75,0000 |
64,7059 |
85,2941 |
66,6667 |
62,5000 |
|
|
32 |
72,9167 |
58,8235 |
85,2941 |
54,5455 |
55,5556 |
|
|
64 |
54,1667 |
47,0588 |
77,4194 |
33,3333 |
39,0244 |
|
|
128 |
65,2778 |
46,3586 |
73,1183 |
62,9630 |
54,5455 |
|
|
256 |
75,0000 |
58,8235 |
83,8710 |
57,1429 |
57,1429 |
In order to evaluate the AlexNet results in Table 7, the summary information of the MBS and LR parameters in terms of best values and mean values are summarized in Table 9 and Table 10.
Table 9Best Results of the AlexNet Architecture in Batch Size and Learning Rate parameters
|
Epoch |
LR |
MBS |
Accuracy |
Sensitivity |
Specificity |
Precision |
F-Score |
|
150 |
10e-3 |
64 |
93,7500 |
100,0000 |
100,0000 |
88,0000 |
|
|
256 |
92,8571 |
||||||
|
10e-4 |
32 |
95,8333 |
94,1176 |
97,0588 |
94,1176 |
94,1176 |
|
|
10e-5 |
8 |
91,6667 |
84,6154 |
||||
|
16 |
91,6667 |
100,0000 |
100,0000 |
||||
|
64 |
92,8571 |
||||||
|
The best |
95,8333 |
94,1176 |
100 |
100 |
94,1176 |
||
The AlexNet results in Table 9 were evaluated from 2 different perspectives to find the MBS and LR parameters that produced the best values. First of all, the best of all values are found. In this part with 3 metrics (Accuracy, Sensitivity, and F-Score), the best results were obtained with the LR parameter 10e-4. In Specificity and Precision metrics, 10e-3,10e-5 LR parameters were obtained. The best results found were obtained with 32 MBS in parameter 10e-4 LR. The other two best results were obtained with 64 MBS in parameter 10e-3 LR and 16 MBS in parameter 10e-5 LR. It is seen that 32 MBS parameter values are optimum in terms of best results. Second, the mean values produced by each LR result are given in Table 10 When these results are examined, it is seen that the best results are obtained in the 10e-4 LR parameter in mean values in 3 metrics (Accuracy, Sensitivity, and F-Score). Therefore, it can be said that 32 MBS and 10e-4 LR parameters are found to be optimum for the Alexnet architecture.
Table 10Mean results of the AlexNet architecture in the MBS and LR parameters
|
Epoch |
Learning Rate |
Mean Accuracy |
Mean Sensitivity |
Mean Specificity |
Mean Precision |
Mean F-Score |
|
150 |
10e-3 |
69,6180 |
53,6414 |
77,3087 |
53,1830 |
61,8912 |
|
10e-4 |
85,1563 |
84,3487 |
89,1208 |
78,3369 |
76,3312 |
|
|
10e-5 |
83,5938 |
78,0462 |
93,8330 |
81,9966 |
71,3516 |
In order to evaluate the GoogleNet results in Table 8, summary information on MBS and LR parameters in terms of best values and mean values is given in Table 11 and Table 12.
Table 11Best Results of GoogleNet Architecture in Batch Size and Learning Rate parameters
|
Epoch |
Learning Rate |
Batch Size |
Accuracy |
Sensitivity |
Specificity |
Precision |
F-Score |
|
150 |
10e-3 |
4 |
94,1176 |
88,8889 |
|||
|
8 |
91,6667 |
86,6667 |
|||||
|
128 |
97,0588 |
||||||
|
10e-4 |
32 |
85,4167 |
82,3529 |
94,1176 |
81,8182 |
||
|
64 |
77,7778 |
||||||
|
10e-5 |
4 |
85,4167 |
82,3529 |
94,1176 |
81,8182 |
75,6757 |
|
|
The best |
91,6667 |
94,1176 |
97,0588 |
88,8889 |
86,6667 |
||
In the GoogleNet results in Table 11, the MBS and LR parameters that produce the best values were found as in the AlexNet assessment. First of all, the best of all values were found. The best results in all metrics were obtained with the 10e-3 learning rate parameter. Two of them were obtained when the batch size was 4, the other two were obtained when the batch size was 8, and one was obtained when the batch size was 128. There is no clear distinction in batch sizes in terms of maximum results. Second, the mean values produced by each learning rate result are given in Table 12.
Table 12Mean results of the GoogleNet Architecture in Batch Size and Learning Rate parameters
|
Epoch |
Learning Rate |
Mean Accuracy |
Mean Sensitivity |
Mean Specificity |
Mean Precision |
Mean F-Score |
|
150 |
10e-3 |
86,1979 |
84,4013 |
91,2081 |
81,1773 |
77,9325 |
|
10e-4 |
82,0313 |
75,2101 |
89,9589 |
73,9771 |
71,4652 |
|
|
10e-5 |
71,9618 |
60,9419 |
83,9382 |
60,6015 |
57,3937 |
When these results are examined, it is seen that the best results are obtained in the 10e-3 LR parameter in all metrics in mean values. Therefore, it can be said that the 10e-3 LR parameters are found to be optimum, while nothing can be said clearly in terms of MBS for the GoogleNet architecture.
At this stage, when both models are evaluated together, LR 10e-4 was determined to be used in all models, since it was obtained with 3 values and a maximum of 10e-4 in terms of LR. Since these 3 best values were obtained with 32 MBS, and there was no MBS that produced more best values when both models were taken together, 32 MBS parameters were accepted as optimum.
3.2. Performance Tests of Architectures Using Optimal Parameters
After parameter optimization, all other models were run based on these two parameters. The final parameters used for all models are given in Table 13 collectively. The results obtained by running the architectures are given in
Table 14.
Table 13Optimized parameter values used in architectures
|
Maximum epoch (ME) |
150 |
|
Initial learning rate (LR) |
10e-4 |
|
Drop factor (DF) |
0.5 |
|
Drop period (DP) |
5 |
|
Minibatch size (MBS) |
32 |
|
Optimizer |
Stochastic Gradient Descent with Momentum (SGDM) |
Table 14Deep learning architecture results obtained withMBS=32 LR=10e-4 parameters
|
Id |
CNN Architectures |
Accuracy |
Sensitivity |
Specificity |
Precision |
F-Score |
|
1 |
googlenet |
85,4167 |
82,3529 |
94,1176 |
81,8182 |
75,6757 |
|
2 |
densenet201 |
79,1667 |
76,4706 |
85,2941 |
68,4211 |
72,2222 |
|
3 |
mobilenetv2 |
79,1667 |
70,5882 |
97,0588 |
83,3333 |
64,8649 |
|
4 |
resnet18 |
75,0000 |
70,5882 |
82,3529 |
63,1579 |
66,6667 |
|
5 |
resnet50 |
83,3333 |
82,3529 |
85,2941 |
73,6842 |
77,7778 |
|
6 |
resnet101 |
72,9167 |
58,8235 |
82,3529 |
55,5556 |
57,1429 |
|
7 |
shufflenet |
68,7500 |
58,8235 |
76,4706 |
52,6316 |
55,5556 |
|
8 |
nasnetmobile |
64,5833 |
47,0588 |
76,4706 |
50,0000 |
48,4848 |
|
9 |
efficientnetb0 |
66,6667 |
58,8235 |
79,4118 |
47,6190 |
52,6316 |
|
10 |
vgg16 |
89,5833 |
76,4706 |
97,0588 |
90,9091 |
80,0000 |
|
11 |
vgg19 |
77,0833 |
82,3529 |
91,1765 |
66,6667 |
70,0000 |
|
12 |
squeezenet |
79,1667 |
52,9412 |
91,1765 |
70,0000 |
58,3333 |
|
13 |
alexnet |
95,8333 |
94,1176 |
97,0588 |
94,1176 |
94,1176 |
|
14 |
darknet19 |
81,2500 |
76,4706 |
94,1176 |
77,7778 |
72,2222 |
|
15 |
darknet53 |
72,9167 |
70,5882 |
91,1765 |
61,1111 |
62,8571 |
|
16 |
inceptionv3 |
75,0000 |
52,9412 |
91,1765 |
62,5000 |
48,6486 |
|
17 |
xception |
68,7500 |
64,7059 |
88,2353 |
55,0000 |
59,4595 |
|
18 |
inceptionresnetv2 |
64,5833 |
64,7059 |
85,2941 |
46,6667 |
51,1628 |
|
19 |
nasnetlarge |
79,1667 |
76,4706 |
91,1765 |
70,0000 |
63,4146 |
When the results in
Table 14 are examined, it is seen that the best values in all metrics are obtained with the AlexNet architecture. In the specificity metric, the best value was obtained by mobilenetv2 and vgg16 architectures along with AlexNet. The second best results were obtained in the vgg16 architecture in Accuracy, Precision and F-Score metrics.
Finally, according to the randomly determined parameters, the results of the pretests given in Table 5.
The final results in
Table 14 were compared on the basis of architectures, and how they performed according to the optimized parameters was examined. The results are given in Table 15.
When the results in Table 15 are evaluated, an increase was observed in 64 out of a total of 95 (19X5) metrics. In total, 15 architectures (Googlenet, densenet201, mobilenetv2, resnet50, shufflenet, vgg16, vgg19, squeezenet, alexnet, darknet19, darknet53, inceptionv3, xception, inceptionresnetv2, and nasnetlarge) displayed an increase in results. When results from resnet18, resnet101, nasnetmobile, and efficientnetb0 architectures are trained with optimized parameters, they performed poorly with reduced results.
3.3. Comparison of the architectures with the literature results
In order to make a fair comparison in the literature, a standard database containing wound images was not used. The studies conducted tests on the image dataset they collected. In particular, there is no study on which diabetic foot ulcer and Pressure ulcer pictures belong to granule, necrotic and slough classes. Studies have generally inferred on a binary classification that shows whether the wound images considered in the relevant study belong to that type or not, rather than tissue classification of the wound image. On the other hand, Rostami et al. studied which wounds belonged to venous, diabetic, pressure, and surgical classes [11]. The success of their study's 3-class Classification is discussed comparatively in the literature.
In Table 16, studies that classify similar or different types of wound images are discussed in comparisons since they are similar in terms of image properties.
Table 16 Literature comparison
|
|
Accuracy |
Sensitivity |
Specificity |
Precision |
F-Score |
|
[10] |
82 |
93 |
75 |
|
83 |
|
[11] |
87.7 |
87.3 |
|
87.98 |
87.44 |
|
[12] |
|
93.6 |
|
95.4 |
94.5 |
|
[13] |
92.5 |
93.4 |
91.1 |
94.5 |
93.9 |
|
[14] 1st row for ischaemia 2nd row for infection |
90.3 |
88.6 |
92.1 |
91.8 |
90.2 |
|
72.7 |
70.9 |
74.4 |
73.5 |
72.2 |
|
|
[15] |
96.4 |
98.4 |
95.1 |
92.6 |
95.4 |
|
|
|
|
|
|
|
|
Alexnet (ours) |
95.83 |
94.11 |
97.05 |
94.11 |
94.11 |
|
Vgg16 (ours) |
89.58 |
76.47 |
97.05 |
90.90 |
80.00 |
In Table 16, the results of the Alexnet architecture, which achieved the best results from the CNN architectures discussed in this study, and the vgg16 architectures, which produced the best results in Accuracy, specificity, precision and F-Score metrics, were compared with the results of other studies in the literature that classified similar wound images. In the specificity metric, the results of the Alexnet and Vgg16 architectures used in this study are better than all the methods. Of all the studies examined in terms of Accuracy and Sensitivity metrics, the Alexnet architecture, which yielded the second-best results, produced competitive results compared to the results in the study by Das et al. [15]. Its results in the precision metric (measure) were very close to the results reported in the studies by Alzubaidi et al. [12] and Goyal et al.[13]. Finally in the F-score metric, it proved to be competitive, producing values very close to those in the studies by [15] [12].
Finally, calculation of ROC Curves of 19 architectures in this study is shown in Figure 4, and calculation of the ROC curves of AlexNet architecture, from which the best values were obtained in this study, is shown in Figure 5.
Diagnosis, care and treatment of pressure ulcers incur exorbitant costs for healthcare systems. Evaluation of the wound, accuracy of diagnosis and successful treatment processes are of critical importance for the patient. In this study, identification of deep learning approaches that provide automatic classification of wounds with high accuracy is discussed. It is known that the parameters affect the success of the architectures. For this reason, the effects of parameters used in architectures are also discussed.
In this study, the success of 19 CNN architectures in the classification of diabetic and pressure ulcer images was examined. The effects of the training parameters of the CNN architectures were also discussed in the examination of the success of the images in categories of granule, necrotic, and slough.
In parameter optimization, experiments were first conducted using randomly selected 10 epoch, 32 batch size and 10e − 4 learning rate parameters. Tests with different epoch numbers were carried out on AlexNet and GoogleNet architectures, which yielded good results in these experiment. Different LR and MBS parameters were tried and interpreted with 150 ME, in which the best values were produced. In the final stage, experiments were carried out by using LR = 10e − 4 and MBS = 32 parameters, which produce the best values in these two parameters, in the 19 CNN architectures specified in the study. In these experiments, AlexNet also achieved the best results on all metrics. Besides AlexNet, mobilenetv2 and vgg16 architectures achieved the best values in the Specificity metric. The second-best results were achieved in the vgg16 architecture in Accuracy, precision, and F-Score metrics.
When the results are evaluated in terms of googleNet architecture, which was in the second place in finding the optimum parameter values in the tests, it is seen that it is in the third place after alexnet and vgg16 in the Accuracy metric. In terms of the Sensitivity metric, vgg16 has the second-best value, along with resnet50. Alexnet, which produces the joint best value in the specificity metric, has the second-best value after vgg16 and mobilenetv2. On the other hand, in Precision and F-Score metrics, Alexnet ranks fourth after vgg16 and resnet50 architectures.
When the metric results produced in all architectures are evaluated, it is seen that the results of the nasnetmobile and efficientb0 architectures are poor. Comparing the posttest results to the pretest results, inceptionresnetv2 and vgg16 architectures are the architectures that increased their metric values the most in total. Likewise, resnet101 and efficientnetb0 architectures were the architectures whose metric values decreased the most.
In this paper, the accuracy of the architectures trained on real data has been researched and it has been demonstrated in the experiments that Alexnet architecture has achieved a high rate of classification success in terms of wound images. Further studies can aim at achieving higher classification success by making improvements in the architectures that are successful in the dataset and via parameter optimization. In addition, the applicability of deep learning methods on different types of wound images can be researched.
Data Availability
The datasets generated and analyzed during the current study are not available to the public due to the fact that they belong to patients who are still on treatment and also because academic studies have not yet been completed, but are available from the corresponding author upon reasonable request.
Data Deposition Information Including Weblink
Data from the study is deposited at: The Unit for Vascular Surgery Research Department, Karamanoğlu Mehmetbey University, Karaman Education and Research Hospital, Karaman, Turkey. https://karamaneah.saglik.gov.tr/
For more requests: http://blog.kmu.edu.tr/huseyineldem/2022/04/14/about-wound-database/
Funding
The authors did not receive support from any organization for the submitted work.
- Haesler E (2014) Prevention and treatment of pressure ulcers: clinical practice guideline. Cambridge media
- Panel NPUA (2009) Prevention and treatment of pressure ulcers: clinical practice guideline. Cambridge Media
- Pieper B (2012) Pressure ulcers: impact, etiology, and classification. Acute and Chronic Wounds: Current Management Concepts, 4th edn. Mosby, St Louis, MO, pp 123–135
- Sen CK et al (2009) Human skin wounds: a major and snowballing threat to public health and the economy. Wound repair and regeneration 17(6):763–771
- Jodheea-Jutton A, Hindocha S, Bhaw-Luximon A (2022) Health economics of diabetic foot ulcer and recent trends to accelerate treatment. The Foot, : p. 101909
- Ogurtsova K et al (2022) IDF diabetes Atlas: Global estimates of undiagnosed diabetes in adults for 2021. Diabetes research and clinical practice, 183: p. 109118
- Sun H et al (2022) IDF Diabetes Atlas: Global, regional and country-level diabetes prevalence estimates for 2021 and projections for 2045. Diabetes research and clinical practice, 183: p. 109119
- Powers JG et al (2016) Wound healing and treating wounds: Chronic wound care and management. J Am Acad Dermatol 74(4):607–625
- Tripathy S, Singh R (2022) Convolutional Neural Network: An Overview and Application in Image Classification. in Proceedings of Third International Conference on Sustainable Computing. Springer
- Shenoy VN et al (2018) IEEE International Conference on Bioinformatics and Biomedicine (BIBM). 2018. IEEE
- Rostami B et al (2021) Multiclass Wound Image Classification using an Ensemble Deep CNN-based Classifier. Computers in Biology and Medicine, : p. 104536
- Alzubaidi L et al (2020) DFU_QUTNet: diabetic foot ulcer classification using novel deep convolutional neural network. Multimedia Tools and Applications 79(21):15655–15677
- Goyal M et al (2018) Dfunet: Convolutional neural networks for diabetic foot ulcer classification. IEEE Trans Emerg Top Comput Intell 4(5):728–739
- Goyal M et al (2020) Recognition of ischaemia and infection in diabetic foot ulcers: Dataset and techniques. Comput Biol Med 117:103616
- Das SK, Roy P, Mishra AK (2021) DFU_SPNet: A stacked parallel convolution layers based CNN to improve Diabetic Foot Ulcer classification. ICT Express
- Szegedy C et al (2015) Going deeper with convolutions. in Proceedings of the IEEE conference on computer vision and pattern recognition.
- Huang G et al (2017) Densely connected convolutional networks. in Proceedings of the IEEE conference on computer vision and pattern recognition.
- Sandler M et al (2018) Mobilenetv2: Inverted residuals and linear bottlenecks. in Proceedings of the IEEE conference on computer vision and pattern recognition.
- He K et al (2016) Deep residual learning for image recognition. in Proceedings of the IEEE conference on computer vision and pattern recognition.
- Zhang X et al (2018) Shufflenet: An extremely efficient convolutional neural network for mobile devices. in Proceedings of the IEEE conference on computer vision and pattern recognition.
- Zoph B et al (2018) Learning transferable architectures for scalable image recognition. in Proceedings of the IEEE conference on computer vision and pattern recognition.
- Tan M, Le Q (2019) Efficientnet: Rethinking model scaling for convolutional neural networks. in International conference on machine learning. PMLR
- Simonyan K, Zisserman A (2014) Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556,
- Iandola FN et al SqueezeNet: AlexNet-level accuracy with 50x fewer parameters and < 0.5 MB model size. arXiv preprint arXiv:1602.07360, 2016.
- Krizhevsky A, Sutskever I, Hinton GE (2012) Imagenet classification with deep convolutional neural networks.Advances in neural information processing systems, 25
- Sunitha Nandhini A et al (2022) Pancreases Segmentation and Classification Based on RCNN and AlexNet. in International Conference on Computing, Communication, Electrical and Biomedical Systems. Springer
- Redmon J, Farhadi A (2017) YOLO9000: better, faster, stronger. in Proceedings of the IEEE conference on computer vision and pattern recognition.
- Redmon J, Farhadi A (2018) Yolov3: An incremental improvement. arXiv preprint arXiv:1804.02767,
- Szegedy C et al (2016) Rethinking the inception architecture for computer vision. in Proceedings of the IEEE conference on computer vision and pattern recognition.
- Chollet F (2017) Xception: Deep learning with depthwise separable convolutions. in Proceedings of the IEEE conference on computer vision and pattern recognition.
- Szegedy C et al (2017) Inception-v4, inception-resnet and the impact of residual connections on learning. in Thirty-first AAAI conference on artificial intelligence.
- Fawcett T (2006) An introduction to ROC analysis. Pattern Recognit Lett 27(8):861–874
No competing interests reported.