-At the present day, the problem is subsisting associated with a reliable diagnosis as soon as possible, especially in medicine in cases of diagnosis of neoplasms. The article discusses research method for the diagnosis of brain diseases in magnetic-resonance tomography images, based on deep learning. This paper presents a novel approach to solutions pattern classification, was formed the optimal architecture convolutional neural network. As a result of experimental studies, was undertake a study major characteristic of convolutional neural network. Through the use of this neural network architecture 95 % the images from the validation set were classified correctly. In addition, the results can be used as an intermediate result for further images analysis.
I. INTRODUCTION
There is a problem in medicine to make a reliable
diagnosis in the shortest possible time due to the growing
volume of medical research [
An algorithm for solving a similar problem using the
classifier was presented in [
This article was based on the idea of achieving higher classification rates on the same set of inputs, but on the basis of a different, more modern method. To show the effectiveness of this method over an algorithm based on an ensemble of decisive trees, this study will use the following quality metrics: – precision – recall – F1-score – specificity – accuracy J P N TP / N TP N FP
J R N TP / N TP N FN F1 2 J P J R / J P J
R
J S N TN / N TN N FP where is N TP – number of correctly classified images with
J A N TP N TN / N TP N TN N FP N FN pathological changes, N TN – number of correctly classified images without pathological changes, N FP – number of images without any pathological changes classified as they have one, and N FN – number of images with pathological changes classified as they have no pathological changes.
The sources of the research materials were the archives
CENTRAL, supported Washington University School of
Medicine, on the platform XNAT [
J 0.81
Currently, the common method for solving classification
problems is the use of convolutional neural networks. They
were first described in the last century [
J
MRI with neoplasms mistake 0.17 0.00 0.10 0.29 0.12
As shown in Fig. 1, initially the network consists of 5 convolutional layers, subsampling layers and fully connected layers. Input images were downscaled from 512 × 512 reports to 72 × 72 pixels. Training lasted for 120 epochs. On convolution and dense layers (except for the last dense layer), the ReLu activation function was used:
f R eLu x m a x x , 0 .
To estimate losses, we used binary cross-entropy (2), where N is the number of images in the sample, yi - is the class of the i-th image, pi – is the output of the neural network for the i-th image.
H p 1 N
( y i lo g ( p i ) (1 y i ) lo g (1 p i ))
N i 1
Table II shows the performance metrics (1) for the convolutional neural network, which was designated as the original. Compared to Table I, the F1-score is 3% higher, however, this change does not affect the quality of classification so significantly.
To evaluate the performance indicators, we also used the
ROC - curve [
The descriptive power of evaluating the effectiveness of
classification using Precision-Recall curve was discussed in
[
A P R n R n 1 Pn
n where Rn and Pn precision and recall on threshold n.
III. RESEARCH OF EFFICIENCY OF THE CLASSIFICATION ALGORITHM BASED ON A
CONVOLUTIONAL NEURAL NETWORK To improve quality assessments (1), it was decided to consistently research the effect of various characteristics of the convolutional neural network architecture on the efficiency. The first research was to determine the effect of the size of the input image. As we have already said in Section 2, the original image was 512 × 512 pixels, and was downscaled to 72 × 72 pixels. To avoid significant changes in the speed of the convolutional neural network, the research was performed on images of the brain MRI of the following sizes: 32 × 32, 40 × 40, 48 × 48, 56 × 56, 64 × 64, 72 × 72, 84 × 84, 92 × 92. It should be emphasized the original image size remains the same, 512 × 512 samples, directly resized input image.
Figure 3 shows a graph of the dependence of the size of the input image on the F1-score and accuracy. As we can see at the graph, it takes the greatest value at the sizes of the input image in 84 × 84 and 92 × 92 pixels. The size of 92 × 92 pixels was chosen as the most favorable size for the input image in solving this problem.
The quality assessment of the algorithm with changing the size of the input image is presented in table II. Fig 3. Dependence of the F1-score and the accuracy on changes in input image size parameters for a convolutional neural network. J A 0.89 0.96 0.92 0.86 0.92 0.17 0.00 0.10 0.29 0.12 B. Effect of changes in convolutional layers on algorithm efficiency
The next stage of the work was the study of the effect of convolutional layers of the neural network on the efficiency of the algorithm.
Figure 4 shows the F1-score and the accuracy depending of the number of the study. On the x-axis the numbers of experiments (from 1 to 5) are shown. In the experiment No. 1, the activation function of the second and the third layer was replaced to sigmoid. In the experiments No. 2 and No. 3, the number of filters in the third layer was increased from 384 to 512 and the number of filters in the second layer was increased from 256 to 384, respectively. An additional sixth convolutional layer (between the fourth and fifth layer) was also added in the experiment No. 5. This layer had a convolution kernel 3 × 3 and 256 filters. In the experiment No. 4, on the contrary, there were only four convolutional layers. The layer between the third and fifth convolutional layers has been removed from the original architecture.
As shown in the Fig. 4, these experiments did not bring
positive dynamics in improving the quality of classification,
but rather worsened the achieved result. Moreover, in the
experiment No. 5, overfitting occurred [
Fig 4. The effect of changes in convolutional layers on the efficiency of the algorithm.
C. The influence of the numerical values of convolution kernels on the efficiency of the algorithm
The final stage in the work was the research of the influence of the numerical values of the convolutional kernels of the convolutional network layers on the classification results. Initially, the second layer, according to the original architecture of the convolutional neural network, had a kernel of 11 × 11 samples. As experiments, convolution kernels of 9 × 9, 7 × 7, 13 × 13 samples were taken. The results of the research are shown in table IV. Fig 5. The effectiveness of neural network training depending on the era number (train_acc and val_acc are the confidence values (1) for the training and for the control sample, train_loss and val_loss are the values of the loss function (2) for the training and for the control sample).
By so doing, as can be seen from the table, the convolution kernel of 9 × 9 samples is the most effective for solving the considered problem for this convolutional neural network architecture. To estimate the losses of the final model, binary cross–entropy (2) was used. The assessment was made after each era, this can be seen in Fig. 5.
KARNEL kernel 9 × 9 J mistake kernel 7 × 7 J mistake
CONCLUSION
The article examined various methods for improving the classification of brain MRI images based on a convolutional neural network. An original convolutional neural network architecture was developed that provides maximum efficiency for solving this problem.
In the course of experimental studies on a set of MRI images, the average quality metrics of classification algorithm became higher than the metrics of another method based on an ensemble of decision trees. At the same time, the need for expencive preprocessing of the input data has disappeared, which may give an increase in the speed of processing MRI images when applied in practice. The value of the F1-score was 95%, and the probability of erroneous classification was equal to 6%. This shows that the classification accuracy is higher than the method described in section 1. By so doing, we can recommend using the resulting convolutional neural network architecture for the recognition of pathologies in images of brain MRI.
These research results can be used to create a computer system for the diagnostics of various pathologies from MRI images of the human brain.
ACKNOWLEDGMENT
The work was partially funded by the Russian Foundation for Basic Research under grants No. 19-2901235 and 19-29-01135 (theoretical results) and the RF Ministry of Science and Higher Education within the government project of the FSRC “Crystallography and Photonics” RAS under grant No. 007-GZ/Ch3363/26 (numerical calculations).