Deep Learning Based Melanoma Diagnosis Identification

VGGNet is a convolutional neural network model proposed by Simonyan [11] and Zisserman. VGGNet explores the relationship between the depth of a convolutional neural network and its performance, and demonstrates that increasing the depth of the network can affect the final performance of the network to some extent. VGGNet contains a total of six network models, which are similar in structure, with the difference lies in the number of sub-layers in each convolutional layer, and the total network depth ranges from11 to 19 layers. Under a single test scale [11], the size of the test image is set as shown in equation (1): (1) Q=S,for fixed S Q = S,for\;{\rm{fixed}}\;{\rm{S}}

The size when the test image is jittered set as shown in the equation (2): (2) Q=12(Smin+Smax), for S∈[Smin,Smax] Q = {1 \over 2}\left( {{S_{\min }} + {S_{\max }}} \right),\;for\;S \in \left[ {{S_{\min }},{S_{\max }}} \right]

Where Q is the test set image, S is the training set image. The top-1 error and the top-5 error are 27.0% and 8.8% respectively for VGG16 with smallest image side=256. The top-1 error and the top-5error for smallest image side = [256; 512] are 25.6% and 8.1%, respectively. The error rate decreases the most significantly and the combined error rate is the lowest, which shows that VGG16 is the best model in VGGNet.

VGG16 consists of 5 convolutional layers, 3 fully connected layers and softmax output layers. Each convolutional layer is followed by one max-pooling layer, and the ReLU function is used for the activation units of all hidden layers. The ReLU function is shown in Figure 2.

Figure 2.

The expression of the ReLU function is shown in equation (3) (4): (3) ReLU(x)=max(x,0)={0,if x<0x,if x>0 {\rm{\;Re\;}}LU\left( x \right) = {\rm{\;max\;}}\left( {x,0} \right) = \left\{ {\matrix{ {0,if\;{\rm{x}} < 0} \hfill \cr {x,if\;{\rm{x}} > 0} \hfill \cr } } \right. (4) {if x<0,f(x)=0,f′(x)=0if x>0,f(x)≈x,f′(x)=1 \left\{ {\matrix{ {if\;x < 0,f\left( x \right) = 0,f^\prime\left( x \right) = 0} \hfill \cr {if\;x > 0,f\left( x \right) \approx x,f^\prime\left( x \right) = 1} \hfill \cr } } \right.

The ReLU function adjusts any number less than zero to zero, which makes the network partly sparse, thus reducing the overdependence between parameters and alleviating the problem of network overflow.

VGG16 uses multiple convolutional layers with smaller convolutional kernels instead of one convolutional layer with a larger convolutional kernel. The perceptual field size obtained from a stack of two 3×3 convolutions is equivalent to a 5×5 convolution, As shown in Figure 3. while the perceptual field size obtained from a stack of three 3×3 convolutions is equivalent to a 7×7 convolution. We illustrate the principle of this substitutability with an example, and the results are shown in Table 1.

Figure 3.

The three Fully-Connected (FC) layers: the first two have 4096 channels each; the third performs 1000-way ILSVRC classification and thus contains 1000 channels. And Max-pooling is performed over a 2×2 pixel window; with stride 2(The effect is to halve the image size). In summary, we can obtain the network structure diagram of VGG16, as shown in Figure 4. From it, we can see that the input of VGG16 network is a fixed size 224×224 RGB image. The only preprocessing is subtracting the mean RGB value, computed on the training set, from each pixel.

Figure 4.

The VGG16 network was trained on the publicly available International Skin Imaging Collaboration (ISIC-2016) training dataset[12], which was selected for testing since the organizers provided real-world labels in both the ISIC-2016 and ISIC-2017 test datasets. The images in the ISIC-2016 dataset were 8-bit RBG with the resolution is 540 × 722 ~ 4499 × 6748 pixels, and since the training and test datasets of ISIC-2016 contain 900 and 379 images, respectively, and the proportion of malignant images in the training and test sets is 19.2% and 19.8%, respectively, considering that malignant images account for a relatively small percentage and will have an impact on the network classification effect, choosing 374 malignant tumors images from the ISIC-2017 dataset were added to the training and test datasets of ISIC-2016, in which 300 images were added to the training set and 74 images were added to the test set.

The network was implemented in the keras framework using the python programming language with a Tensorflow2-GPU backend, and the experiments were conducted on a windows 11 operating system with the following hardware configurations: AMD RYZEN5 4000 series CPU @ 3.60 GHz × 16 processor, GeForce GTX1660TI GPU with 8GB GDDR5 memory.

This study is a dichotomous classification of images, in the dichotomous case; the model finally needs to predict the outcome in only two cases. For each category our predictions are obtained with probabilities P and 1-P, when the expression of the cross-entropy loss function as shown as equation (5). (5) L=1N∑i−[yilogpi+(1−yi)log(1−pi)] L = {1 \over N}\sum\limits_i { - \left[ {{y_i}\log {p_i} + \left( {1 - {y_i}} \right)\log \left( {1 - {p_i}} \right)} \right]}

Where y_i denotes the label of sample i, positive class is 1, negative class is 0. p_i denotes the probability that sample i is predicted to be a positive class.

Comparisons were made using the most common skin lesion segmentation evaluation metrics, including Accuracy, Precision, Sensitivity, Recall, and Specificity. These metrics were used in the evaluation of the mode. They are illustrated below:

Accuracy: It measures the proportion of true results (both true positives and true negatives) among the total number of cases examined. Accuracy is expressed as: (6) Accuracy=TP+TNTP+TN+FP+FN Accuracy = {{TP + TN} \over {TP + TN + FP + FN}}

Precision: It measures the accuracy and represents the proportion of examples classified as positive that are actually positive. Precision is expressed as: (7) Precision=TPTP+FP Precision = {{TP} \over {TP + FP}}

Sensitivity: It measures the proportion of those with positive values among those who are actually positive. Sensitivity is expressed as: (8) Sensitivity=Recall=TPTP+FN Sensitivity = Recall = {{TP} \over {TP + FN}}

Specificity: This is the proportion of those that are negative among those who actually tested negative. Specificity is expressed as: (9) Specifivity=TNTN+FP Specifivity = {{TN} \over {TN + FP}}

Eval_Top1: This is the label of the network prediction that takes the largest one inside the final probability vector as the prediction result, and if the one with the largest probability in the prediction result is correctly classified, the prediction is correct. Eval_Top1 is expressed as: (10) Eval_Top1=2⋅Precision⋅RecallPrecision+Recall Eval\_Top1 = {{2 \cdot Precision \cdot Recall} \over {Precision + Recall}}

Eval_Top5 is the top five with the largest probability vector at the end, as long as the correct probability occurs, the prediction is correct. Eval_Top5 is expressed as: (11) Eval_Top5=(1+52⋅Precision⋅Recall)52⋅Precision+Recall Eval\_Top5 = {{\left( {1 + {5^2} \cdot Precision \cdot Recall} \right)} \over {{5^2} \cdot Precision + Recall}}

Where FP is the number of false positive pixels, FN is the number of false negative pixels, TP is the number of true positive pixels and TN is the number of true negative pixels.

In this study, the network is evaluated using a confusion matrix, where each column of the confusion matrix expresses the category prediction of the classifier for the sample, and each row of the matrix expresses the true category to which the sample belongs. As shown in Table 2.

The following three equations can be obtained from the Table 2: (12) Precision=aa+c Precision = {a \over {a + c}} (13) Recall=aa+b Recall = {a \over {a + b}} (14) Accuracy=a+da+b+c+d Accuracy = {{a + d} \over {a + b + c + d}}

Precision, Recall and other parameters calculate the characteristics of a certain classification, while Accuracy is a criterion to determine the overall classification model.