Medical image processing technologies allow for automating and improving diagnostic and analysis processes, providing doctors with more accurate and faster results. The use of artificial intelligence, deep learning, and computer vision allows for the creation of efficient and automated systems that can detect pathologies, classify images, and provide valuable decision support to doctors. The description and preliminary processing of the data set, which is a key stage for the preparation of system input data, has been performed. Models for training are also developed, including the selection and tuning of neural network architectures. The introduction of a new method for training a neural network turned out to be very successful. This approach significantly improved the training quality of the model, helping to increase the accuracy and ability of image classification. The application of this method significantly improved the efficiency and reliability of the X-ray image recognition system. The research results indicate that the new learning method, based on the combination of Adam and SGD methods, raised the accuracy of image recognition to the level of 95-97% while increasing the training time by only 1-2%. The developed system can be considered as an initial version that paves the way for further improvement. It was determined that the main driving factor for improving the system is the developed neural network training method.
The topic of creating an information system for processing and analyzing medical images using artificial intelligence and developing a system for processing medical images in real time are extremely relevant and important in the field of medicine [1]. Here are some aspects of the relevance of these topics: • Improvement of diagnostic accuracy.
Using artificial intelligence to analyze medical images can help doctors detect symptoms and pathologies that may be difficult to detect using traditional methods. This can lead to more accurate and earlier diagnoses. • Reducing the burden on medical personnel. Medical imaging systems can automate the analysis process, helping doctors focus on important tasks and reducing the burden on medical staff. • Quick access to information. The development of systems for processing medical images in real-time will allow doctors to have instant access to the results of the analysis, which is especially useful in urgent situations, such as injuries or serious diseases [2]. • Reducing the risk of errors. Artificial intelligence can help weed out false positives and increase the reliability of The purpose and objectives of the research. diagnoses [3]. The main goal of the development is to create • Expanding access to medical care. Such a system for processing and analyzing medical systems can help reduce inequalities in images aimed at improving X-ray diagnostics access to health care, as they can be used of chest organs. This system is designed to in different health facilities, including increase the accuracy and efficiency of remote areas [4]. detection of pathologies and diseases, as well • Discovery of new research opportunities. as to optimize the time and resources of the Processing and analysis of medical medical staff. images can help in the development of new research and approaches to the 2. Statement of Research Problem treatment of various diseases.
In this regard, the development of systems Modern technologies allow automation of the for processing and analyzing medical images processing and analysis of medical images, using artificial intelligence and systems for which contributes to the efficiency and processing medical images in real-time has accuracy of diagnosis. This is especially great potential for improving the quality of important in the conditions of increasing medical care and saving patients’ lives. volume of medical data.
In the process of developing the system, the The use of innovative technologies in the peculiarities of the medical field and specific processing and analysis of medical images requirements related to the processing and opens up new opportunities for accurate analysis of medical images should be taken into diagnosis, treatment, and monitoring of account. This includes taking into account a patients. Such studies contribute to the high standard of confidentiality and data development of automated systems that help security, as medical information is particularly specialists in the fast and reliable sensitive. It is also important to take into interpretation of medical images. account the variety of types of medical images, In particular, the development and which requires the development of flexible improvement of algorithms for the recognition algorithms capable of working with different of pathologies on X-ray images, detection of formats and modalities. signs of diseases on computer tomography, and
In addition, it is important to consider the analysis of other medical images help to
specifics of interaction with medical personnel, improve the speed and accuracy of diagnosis.
providing a convenient and efficient user Given the rapid pace of technology
interface for interacting with the system. The development and the constant replenishment
possibility of integrating the system with of medical databases, research in this direction
existing medical information systems to ensure is important for ensuring effective and modern
interaction and data exchange should also be medical practice.
taken into account [
In addition, this study may have practical processing helps to improve the accuracy of
value for developers, simplifying their work diagnosis of various diseases. The ability to
and opening new opportunities for the automatically detect pathologies and
implementation of advanced technologies in abnormalities in images allows for early
medical practice [
and already available similar solutions for the
classification of objects in images, it was decided
to develop the technical requirements for the
system that is planned to be developed [
The development of a system for processing and analyzing medical images and classifying objects on them is a complex task and requires the definition of technical requirements for the successful implementation of the project. Here are some general technical requirements that may be important for this process: 1. Processing and storage of medical images. The system must be able to download and store medical images, in particular DICOM images, which are widely used in medicine. 2. Segmentation and definition of regions of interest. The system should be able to automatically identify and highlight Regions of Interest (ROIs) on medical images for further processing and analysis. 3. Using deep learning. It is recommended to use deep learning, in particular Convolutional Neural Networks (CNN), to recognize and classify objects in images. 4. Ability to track and analyze changes in real-time. The system should support the analysis of changes in medical images in real-time, in particular for patient monitoring. 5. Support for various types of medical images. The system should be universal and support various types of medical images, such as X-rays, CT, MRI, etc. 6. Data protection and confidentiality.
Ensuring a high level of security and confidentiality of medical data, including data storage and transfer requirements. 7. Possibility of integration with other systems. The system must be able to integrate with other medical systems, hospital information systems, and data management systems. 8. Training and retraining of models. The ability to train and retrain models to improve the accuracy of medical image classification and analysis. 9. User interface. Development of a convenient user interface for doctors and medical professionals to navigate, visualize, and analyze results. 10. Documentation and Support. Provision of documentation, instructions, and technical support for system users. 11. Compliance with regulatory requirements. Consideration of regulatory requirements and standards in the medical field when developing and operating the system. 12. Evaluation of results and metrics.
Establishing metrics to evaluate system accuracy and performance, such as sensitivity, specificity, accuracy, and others. 13. These requirements can be adapted according to the specific needs of further development of the system.
Scientists and researchers use a variety of
chest X-ray image datasets to study and train
algorithms for the recognition of opacities,
pneumonia, and COVID-19. Among these
datasets, several popular ones can be singled
out, which have become key tools in the study
and development of algorithms for the
automatic detection and classification of lung
diseases [
Below is a list of several important datasets
used in this area:
1. The COVID-19 Image Data Collection is a
dataset of X-ray images to study the impact
of COVID-19 on the lungs. It was created by
collecting medical images from websites
and publications and currently contains
123 frontal radiographs [
dataset of X-ray images for the study of
pneumonia. The dataset is organized into used to classify new chest X-rays as showing
3 folders (train, test, val) and contains signs of pneumonia or not. This can be done in
subfolders for each image category real-time, making it a potentially valuable tool
(Pneumonia/Normal). There are 5863 x- for healthcare providers in the diagnosis and
ray images (JPEG) and 2 categories management of patients with pneumonia. In
(pneumonia/normal) [
Deep learning algorithms can be trained on large
datasets of chest X-rays to recognize patterns and
features that indicate pneumonia. This involves
the use of CNNs, a type of deep learning
architecture that is particularly well-suited to
image recognition tasks [
Once trained, deep learning models can be
The second dataset taken is ChestX-ray14, which contains 112,120 chest X-rays of 30,085 individuals. Of these 112,120 images, 1,431 images had signs of pneumonia. To obtain a balanced data set, 1431 normal X-ray images (marked as “No Results”) were selected from the data set. Thus, the final collected data set contains 1431 pneumonia images and 1431 normal X-ray images. 80% of the data is used Figure 2: Results of using the model for training, generating 2290 images (1145 Sequential_2 pneumonia images and normal images each), A third modeling approach involved the use of 5% of the data is used for validation, transfer learning using VGG-16 with some generating 142 images (215 pneumonia modifications in the last three levels (Fig. 3). images and normal images each), and the last
Here, the Chest Pneumonia X-ray dataset is 15% of the data is used for testing, yielding 430 used to collect Pneumonia X-ray images that (71 images of pneumonia and normal images). take into account images from various open The count plot of the first data set is used to sources and are reviewed regularly. Here, two display pneumonia counts and normal images. datasets are used to train models for To represent all images, Fig. 3 shows that the pneumonia diagnosis. The first dataset X-axis of the training set contains the values 0 consists of 5856 chest X-ray images, of which (corresponding to 1224 normal images) and 1 4273 are pneumonia images and 1583 are (representing 3418 pneumonia images), while normal chest X-ray images [19]. the X-axis of the test set contains the values 0 (representing 278 normal images) and 1 (representing 641 pneumonia images), and the Y-axis of the training set displays the count graph of both pneumonia and normal images (Fig. 4). A total of 80% of the data is used for training, generating 4642 images (3418 pneumonia images and 1224 normal images), 15% of the data is used for testing, generating 919 images (641 pneumonia cases and 278 normal images), and the last 5% of the data is used for validation ( 214 cases of pneumonia and 81 images without pneumonia).
sources for optimal training to demonstrate improved performance on larger datasets. Since only a small dataset is used, this is used to artificially expand the dataset. It also helps to avoid overfitting. A data augmentation approach has often been used and increases the number of images by applying a series of changes while preserving the class labels [20]. Data augmentation is applied to the pneumonia class training images to increase image diversity, which also acts as a dataset regulator.
The training and validation loss plot can provide information about how the model learns and how well it generalizes its knowledge to the validation data. Loss determines how many errors the model makes in forecasting.
Training Loss—training loss reflects how accurately the model predicts the training data during each epoch. Typically, the loss should decrease during training, indicating that the model is learning to identify patterns in the data.
Validation Loss—Validation loss determines how well the model generalizes its skills to data it has not seen during training. A decrease in validation loss indicates that the model effectively recognizes patterns and avoids overfitting. The graph has the form where the epochs (training iterations) are displayed on the X-axis, and the loss value (from 0 to the highest loss value) on the Y-axis is shown in Fig. 5.
A good scenario is that both curves (training and validation loss) will show a decline, indicating the effectiveness of training and generalization. It is important to ensure that the validation loss does not increase, a sign of possible overtraining. The model training schedule in machine learning can include two important metrics: training accuracy and validation accuracy. Both metrics reflect how well the model has learned to recognize patterns in the data during training and how well it generalizes those skills to new data.
Accuracy on the training set (Training Accuracy)—this metric determines the accuracy of the model on the data it used during training. If the model has learned well, the accuracy of the training set will be high. However, it is important to ensure that the model does not overtrain (overfitting) specific examples in the training data and can effectively recognize new data.
Accuracy)—this metric determines the accuracy of the model on data that it has not seen during training. It is an indicator of how well the model generalizes its knowledge to new, previously unseen data. If the accuracy of the validation set also increases during training, this may be an indication that the model is effectively learning and detecting common patterns, rather than simply “remembering” the training data.
The graph has the form where epochs (training iterations) are displayed on the X axis, and accuracy (from 0 to 1 or percent) on the Y axis in Fig. 6. Plots for training and validation accuracy can help determine model performance and identify potential problems such as overtraining or undertraining. A histogram of people’s ages on X-ray images (Fig. 7), presented together with the number of training images, gives a visual impression of the age distribution in the dataset.
The x-axis of the histogram shows the age ranges, and the y-axis shows the number of images corresponding to each range. This graphical representation can indicate the age at which X-ray examinations are more common or the age distribution of patients in the sample to train the model.
Such analysis can be useful for understanding whether there is a diversity of age groups in the training dataset and whether particular age groups should be given special attention when training the model.
generated to visualize the distribution of male and female patients in the training dataset. Two bars are marked on the x-axis, representing the number of images for each gender. The information provided in Fig. 9 is metadata for a specific X-ray image and contains some key attributes [21, 22]: 1. patientid (patient identifier)—a unique patient identifier that can be used to track and analyze medical information. 2. offset—this value may indicate certain parameters or displacements associated with the
X-ray examination. 3. sex—indicates the gender of the
patient. 4. age—shows the age of the patient at
the time of the X-ray. 5. finding—describes the result of the examination. In this case, “COVID19” indicates the detection of the 2019 coronavirus disease. 6. survival—shows whether the
patient survived (“Y”—yes). 7. view—indicates how the X-ray examination was carried out. In this case, “AP Supine” is an anteriorposterior (anteroposterior) view in the supine position. 8. modality—shows what type of examination was used. In this case, “X-ray” is X-ray radiation. 9. date—indicates the date of the X
ray examination. 10. location—shows the place where
the survey was conducted.
11. url (URL)—a link to a source or
additional information. In this case,
a URL is provided where additional
information can be obtained.
12. clinical notes—indicates clinical or
medical information about the
patient’s condition.
13. other notes—other additional notes
or instructions that may be useful
for understanding the context or
other details of the examination.
The generated pie chart provides information
on the number of images on which different
pathogens and conditions are detected (Fig.
10). Each sector of the diagram corresponds to
a specific pathology, and its size indicates the
percentage or number of images where this
pathology was detected [23].
Implementation of a new method of neural The maximum VGG19 has 16 convolutional
network training layers. +3 fully connected layers. In addition, the
The following algorithms were analyzed: VGG network is not accompanied by a pooling
1. Nesterov Accelerated Gradient layer behind each convolutional layer or a total of
2. Adagrad 5 pooling layers distributed under different
3. Adam convolutional levels.
4. Adamax Each convolution layer (Fig. 11) in AlexNet
5. RMSProp contains only one convolution, and the size of
6. Adadelta the convolution kernel is 7×7. In VGGNet, each
7. SGD. convolution layer contains 2 to 4 convolution
The comparison of algorithms made it operations. The convolution kernel size is 3×3,
possible to stop the use of the Adam algorithm. the convolution step size is 1, the pooling
Adam (Adaptive Moment Estimation) is a kernel is 2×2, and the step size is 2. The most
popular optimization algorithm, especially in obvious improvement of VGGNet is to reduce
deep learning, for several reasons: the size of the convolution kernel and increase
1. Learning rate adaptability: Adam uses the number of convolution layers [
Adam often shows good performance in multidimensional parameter spaces, which is usually characteristic of deep neural networks [25]. Figure 11: Each layer of convolution 4. Defaults: A big advantage of Adam is that it has reasonable defaults that usually work To obtain a 224×224 input image, each scaled well in many cases without the need for image is randomly cropped during each SGD extensive tuning. iteration. To improve the dataset, the cropped Choosing a modified optimization strategy image is also randomly flipped horizontally that involves branching at the damping point of and the RGB color shifted. the approach to the extremum is an important A 1×1 convolution kernel is introduced into decision. This approach can have several the VGGNet convolution structure. Without advantages, such as improving convergence, affecting the input and output dimensions, a adapting to the landscape of the loss function, and nonlinear transformation is introduced to avoiding local minima more efficiently [26]. increase the expressive power of the network
The VGG input is set to a 224×244 RGB and reduce the number of calculations [
Each model has been carefully tuned and evaluated to determine which model performs best in object detection on test sets. This experiment provided an opportunity not only to evaluate the effectiveness of each model but also to compare their characteristics to determine the optimal option for further research and applications.
The experiment included several stages: 1. Selection of test sets. Definition of 2200 test sets which were obtained from different sources such as ResNet-32,
DenseNet, PyramidNet, and SENet. 2. Data preparation. Processing and preparation of the received test sets for use in the experiment, including standardization and other necessary operations. 3. Application of models. Using different models such as ResNet-32, DenseNet, PyramidNet, and SENet to classify objects on test sets. 4. Collection of results. Evaluate the performance of each model based on its ability to classify objects on the test sets. 5. Analysis of the results. Comparing the performance of different models to determine which one exhibits the best accuracy or other important characteristics.
The conducted study showed that the Adam optimization algorithm, as a rule, turns out to be more effective in terms of learning speed compared to the Stochastic Gradient Descent (SGD) method. This is because Adam uses adaptive learning rates for each parameter separately, which allows you to approach the optimum from different angles and reduces the probability of getting stuck in local minima.
This approach can lead to faster convergence of
the model during training [
In the initial training stages, Adam [
For example, consider a situation where the
loss function has many local minima. In this
case, Adam may exhibit high speed during the
initial training phase but may become less
efficient when the loss function approaches one
of the local minima. On the other hand, SGD,
having less computational complexity, may
seem simpler and more efficient at later stages
of learning, where it is important to avoid
getting stuck in local minima [
Therefore, the effectiveness of the SGD and Adam methods may vary depending on specific data, model characteristics, and learning parameters. It is recommended to conduct a series of experiments with different optimization methods for a specific task to determine which one works faster and more efficiently under specific conditions.
Gradient Descent) and Adam have their advantages and disadvantages, which are important to consider when choosing a specific task.
SGD requires less memory because it uses only one random subset for each weight update. SGD may be less sensitive to large amounts of noise in the data because it uses random subsampling. Convergence can be slower, especially in deep neural networks, because SGD can get stuck in local extrema.
Adam adaptively adjusts the learning rate for each parameter, making it efficient in multidimensional spaces. It usually turns out to be effective in practice for different tasks and datasets. Adam uses more memory to store additional information, such as exponentially smoothed gradients and their squares.
The average of test results on test images is presented in Table 1.
The combination of SGD and Adam methods leads to the highest accuracy of the model on the test set, but at the same time increases the training time slightly compared to using SGD or Adam alone.
contributes to improving the accuracy of diagnoses, early detection of diseases, and improving the ability of systems to monitor the dynamics of pathologies over time.
The analysis revealed that there is a large and diverse set of methods in the field of pattern recognition in medicine. Such tasks often require the use of non-trivial methodologies, and the development of original models and algorithms. Among the main challenges are the analysis of complex and hybrid images, as well as the improvement of pattern recognition systems.
The analysis of the subject area and existing solutions indicates that image recognition in radiological studies is an actual and promising field of research. Chest X-ray images data set was analyzed.
This dataset includes chest radiographs that can be used to recognize pathologies and detect abnormalities in the images. A detailed review of this dataset allowed us to consider its features and opportunities to improve the performance of our recognition system. Given the diversity of data in this set, we were able to tune our model parameters to more accurately and reliably detect various pathologies in X-ray images.
The description and preliminary processing of the data set, which is a key stage for the preparation of system input data, has been performed. Models for training are also developed, including the selection and tuning of neural network architectures.
The introduction of a new method for training a neural network turned out to be very successful. This approach significantly improved the training quality of the model, helping to increase the accuracy and ability of image classification. The application of this method significantly improved the efficiency and reliability of the X-ray image recognition system. The research results indicate that the new learning method, based on the combination of Adam and SGD methods, raised the accuracy of image recognition to the level of 95–97% while increasing the training time by only 1–2%.
The developed system can be considered as an initial version that paves the way for further improvement. It was determined that the main driving factor for improving the system is the developed neural network training method. This method is based on the efficient use of Adam and GD methods depending on different input parameters.