CC BY
14TECHNICAL SCIENCES
AN INNOVATIVE APPROACH TO THE DEVELOPMENT OF A MEDICAL DIAGNOSTIC SYSTEM
BASED ON ARTIFICIAL NEURAL NETWORKS
Cozac E.
Masters Degree in Computer Science Software Engineer GAN C/O Memery Crystal Llp, London, United Kingdom
ABSTRACT
The article describes an innovative approach to the development of a medical diagnostic system based on artificial neural networks. It is emphasized that artificial neural networks are mathematical models, as well as their software or hardware implementations, built on the principle of organization and functioning of biological neural networks. It is emphasized that currently the use of statistical data processing methods predominates in medical research, and the most common descriptive methods used in traditional statistical research are survival analysis and multidimensional complex analysis, which is classified as discriminant, cluster, factor and correlation analysis. Data extraction methods are a powerful statistical device in the study of data in comparative analytics, which are used to identify hidden patterns and build predictive models. It is noted that existing neural networks are able to work with both numerical data lying in a certain limited range, and non-numerical parameters, such as graphics of various configurations. It is emphasized that both from a scientific and practical point of view, one of the main advantages of using neural networks is their ability to learn with data analysis, establishing complex and hidden connections and subsequent presentation of independent results. It is emphasized that one of the dominant problems in the application of artificial neural network models is the previously unknown architecture of the designed neural network and its degree of complexity, which will be sufficient for the reliability of the obtained result. The scheme of a typical procedure of diagnostics by a doctor and directly stages of medical diagnostics are proposed, the principle of neuronal training with backpropagation approach is described and an innovative backpropagation algorithm is developed. It is emphasized that, like any other tool for diagnosing diseases, the diagnostic system based on artificial neural networks requires user control to ensure clinical efficacy and safety.
Keywords: artificial neural network, machine learning, approach, medical system, automation, development, innovation.
Introduction. The study of diagnosis of diseases plays a vital role for the medical system of Ukraine. Any reason or circumstances that lead to pain, disease, dysfunction, or, ultimately, human death is called a disease. Diseases can affect a person physically and mentally, and this fact is an influential fakery for human life. The causal study of the disease is called a pathological process. The disease consists of signs or symptoms that are interpreted by clinical experts. Diagnosis was defined as a method for detecting a disease based on its features and symptoms for concluding about its pathology. Diagnosis of diseases is at the same time the most complicated process, with a fundamental basis for a medical worker, as well as to conclude, with regard to illness.
To minimize uncertainty in health diagnostics, experts collect empirical data to analyze patient diseases and diagnose. However, not all doctors have expert knowledge in the fields of medicine. There is a need for an automatic diagnostic system that will enable both human knowledge and accuracy of the machine. In order to achieve the most accurate results of the diagnostic process with reduced costs, an appropriate decision support system is required. Classification of diseases depending on different parameters is a complex task for experts-people, but artificial intelligence will help you identify and process such cases. Currently, various techniques of artificial intelligence are used in the field of artificial intelligence for accurate diagnosis of dis-
eases. Intelligent systems are an integral part of computer science, the life of any intelligent system is learning.
Today, there are various methods of artificial intelligence, based on learning, such as deep learning, machine training, etc. Some specific methods of artificial intelligence that are important in medicine named as an intellectual system based on rules, they act as a decision support system. Gradually, intelligent systems in medicine are replaced by automatic methods based on artificial intelligence, which minimizes human influence.
Literature review. In modern science development, relevance acquires work aimed at introducing neural network methods of machine learning in various spheres of human life, in particular in the medical sphere.
T. B. Martynyuk and Ya. V. Sukruk [1] performed an analysis of the characteristics of a neural network approach to medical express diagnostics. The analysis of methods and means of biomedical diagnostics has been conducted by authors showed relevance and prospects of application of neural network technologies. Scientists are offered a neural network classifier based on an improved Hamming network with the formation of discriminatory functions.
O. V. Zavloro disclosed principles of medical diagnostics of malignant skin cancer with the help of artificial neural networks [2]. The author presents the
principles of medical diagnostics of human skin diseases with the help of artificial neural networks. The aspects of the development of artificial intelligence, which can create intelligent systems in various fields of application on the basis of biological approaches to the basis of biological approaches.
In research [3] the author first used artificial neural network to adjust the dynamic errors in the measuring channel of thebio impedance frequency analyzer, which made it possible for three orders of magnitude to expand the bandwidth and somewhat improve accuracy compared to the known formal method of algorithmic correction.
I. Novoseltsev [4] proposed a neural network method for recognizing image parameters on an example of melanoma using PNN and CNN networks based on which a neural network method for controlling the size of the skin formation and recognition of the heterogeneity of the color of the skin changes. It allow increasing the accuracy of controlling the change in size Observed pigmentary neoplasms of the skin for early diagnosis of melanoma and increase the efficiency of the classification of polychromic of pigmentary neoplasms of the skin.
In the dissertation [5], the use of hybrid models of neural networks, which provide an arbitrary combination of multiple architectures or types of networks in order to increase the accuracy of prediction. Such a combination involves learning a hybrid topology as a whole system. The mathematical model of the hybrid neural network classification is formalized.
M. M. Deeper, K.V. Salad and N.A. Tkach [6] carried out a diagnostic expert-medical system using neural networks.
From foreign authors it is worth noting such works asZhou He & Huang Jianjun & Peng Xuemei[7], Abu-ZeidNoha&KashefRasha&BadawyOsama[8], XueQinghan&ChuahMooiChoo [9], SunBQ&ZhaiAM&FengYJ&WangXF & Wang HQ
[10], Chaturvedi Animesh & Tiwari Aruna & Chatur-vedi Shubhangi & Lio Pietro[11],Li, Aiping & Jin Songchang & Zhang Lumin & Jia Yan[12], Moradpour Amin & Karadima Olympia & Alic Ivan & Ragulskis Mykolas & Kienberger Ferry & Kosmas Panagiotis [13], Onisko Agnieszka & Druzdzel Marek[14], Khandoker Ahsan & Begg Rezaul [15], Takata Hideaki & Nogawa Hiroki & Nagata Hiroshi & Gomi Yuichiro & Tanaka Hiroshi[16], Akbari Zohreh & Unland Rainer [17] and others.
However, taking into account the described scientific gains, on the topic, the issue of developing an effective development of a medical diagnostic system based on artificial neural networks remains open and requires detailed processing.
Objectives. The main article aim is to carry out a description of an innovative approach to the development of a medical diagnostic system based on artificial neural networks.
Research results. The work of the modern medical system is focused on fuzzification and de-fuzzifica-tion of patient's data. Since the patient's data is nothing more than physiological indicators, they are subject to noise and uncertainty. Onecannot always trust this patient as a degree or weight as they depend on the quality and accuracy of the measurement units, as well as skills of a specialist. In addition, to establish a reliable diagnosis, there is a large data available.
Diagnosis, as the basis of the medical system, can also be defined as a method for determining what disease is based on symptoms and signs of a person as shown in Fig. 1. Data collected during a medical examination of a patient with medical pathology form the necessary knowledge for diagnosis. Often, at least one diagnostic procedure, such as medical tests, is carried out during this procedure. To put an accurate diagnosis, the doctor carries out a process that includes several stages that will allow him to collect the maximum amount of information.
Figure 1 - Doctor medical diagnostics stages Figure 2 shown a scheme of a typical procedure for diagnosing that a doctor make.
Patients data base
Doctor
Figure 2 - Doctor scheme for a typical diagnostic procedure
Artificial neural networks with reverse distribution use one of the most popular algorithms for teaching neural networks that is called a reverse distribution algorithm. This algorithm trains a given multilayer neural network with a direct bond for a given set of input templates with known classifications. Thus, each unit of the input data set, the network verifies the input reference template. Then the resulting sample is compared with a known result and the error value is calculated. Based on the received error, the weight of the neural network connection is adjusted. The reverse distribution algorithm is based on the delta-learning rule in which the weight adjustment is carried out through the mean square error of the initial response to the sample.
Stages of the algorithm are the next:
1. Initialization of the weight coefficients of the network, the establishment of minimum values of weights from the generator of pseudorandom sequences.
2. Consistent implementation of the first stage with a constant change of weights to reach a minimum error.
Weight value after update
A (o( t) = -T
dE(t) d(ß
Update the weight of the neural network w(t + 1) = u(t) + A (o(t)
Error value
E(t+1)
n
E=2YJ(dj-yj)2
i=i
ge t - iteration number; m - weight of the connection; dj - the desired result; jj- actual exit value; E- error;
h - training speed.
Error Ecan be selected as a function of the mean square error between the actual j output and the desired result d
Figure3 - Realization principle of neuron training with reverse distribution approach
The described reverse distribution algorithm has a number of disadvantages that are based on the implementation speed, as if the initial teaching rate is set to a small value to minimize the overall error, the learning process will slow down. On the other hand, higher education speeds can accelerate the learning process due to the risk of potential oscillations.
To eliminate these disadvantages, it is effective to use the pulse term. Update weight in reverse distribution algorithm with a pulse aperiod A is determined as follows:
dE(t)
Aw(t) = -T —— + aAw(t - 1)
Then relying on a gradual learning strategy can be updated using the following formula: T(t) = T0+AE(t- 1).
dœ
The adaptive training speed can also be adopted to accelerate the convergence of the algorithm. For batch training strategy, the teaching speed can be configured as follows:
( aT(t - 1)if E(t) < E(t - 1) T(t) = {pT(t - 1)if E(t) > gE(t - 1) T(t - 1)in all other cases
whereh(t) - the training speed is on t iteration, anda, p, Rvalue are chosen in the next way a > 1,0 < p < 1,g > 1.
The innovative reverse distribution algorithm is directed to adjust the error that will be transmitted back from the output level to each intermediate level block. The proposed innovative reverse distribution algorithm will improve the productivity of a commonly known reverse distribution algorithm P. The rate of convergence of the learning process increases, in the reverse distribution algorithm error on one output unit is defined as: f0
Straining vector/output vector
training vector/output vector Qtraining vector/output vector )
Where index «training vector» belongs to the ith educational vector, and «output vector» refers to the i-th output block. Thus,
Wt
training vector/output vector
is the desired initial value,
aQtraining vector/output vector — actual °utput and units,
then Straining vector/output vectorwi^ spread back to update the output level and weight of the hidden level. While an error on one output unit in the adjusted innovative reverse distribution algorithm will look like this:
f0 = 1 + pi^training vector/output vector Qtraining vector/output vector^
Straining vector/output vector ^ 1 ^
In the case if
training vector/output vector Qtraining vector/output vector ) — 0
P0 = _ (l + p(^training vector/output vector —Qtraining vector/output vector^ )
Straining vector/output vector I-1- 1 ° J
In the case if
where^,0
(Wt
training vector/output vector
viewed as
training vector/output vector
new, proposed innovative reverse distribution algorithm. Said algorithm uses two
forms%training vector/output vector , since the function
Initial data input
Qtraining vector/output vector ) <
expalways returns zero or positive values (and the adaptation operation for many outgoing units requires a decrease in actual output data, rather than increase their). Stages of implementation of the proposed algorithm are shown in Figure 4.
> r >
Error calculation of E (t+1)
Calculation of the
error ^(tra|n|ng vector / output vector)
f >
Calculation of incoming values to units of hidden layer
v_
J
Calculation of pure
input values for unit of output level
Figure 4 —Implementation stages of the proposed innovative reverse distribution algorithm
The implementation of the algorithm continues
until the error (^training vector/output vector Qtraining vector/output vector )becomes acceptably
small for each pair of educational vectors.
Conclusion. The paper describes an innovative approach to the development of a medical diagnostic system based on artificial neural networks. When using diagnostic tools based on artificial neural networks in everyday practice of the medical system, they should understand their features of integration into the workflow. The range of diagnostic capabilities based on artificial neural networks is completely determined by the training sample. Like any other diagnostic tool, the diagnostic system based on artificial neural networks requires user control to ensure clinical efficiency and security.
Prospects for further work are the development of a medical diagnostic system using the presented innovative reverse distribution algorithm.
References
1. Martynyuk T. B., Filler J. V. (2019). Neural network approach to medical express diagnostics. Information technologies and computer equipment: Vinnitsa Polytechnic Institute, No. 6. P. 37-44.
2. Zagorodniy O. V. (2020). Principles of medical diagnostics of malignant skin cancer with artificial neural networks / Scientific journal "Computer-Integrated Technologies: Education, Science, Production", Lutsk, 40, P. 31-36.
3. Homa Yu.V. (2020). Theory and methods of computer processing of biosignals based on machine learning. - On the rights of the manuscript. Dissertation for the degree of Doctor of Technical Sciences in specialty 05.13.13 (Computer systems and components), National University "Lviv Polytechnic", MON of Ukraine, Lviv, 379 p.
4. Novoseltsev I.V. (2019). Methods and means of recognizing changes in the properties of an object by image based on artificial neural networks. Qualification scientific work on the rights of the manuscript. Thesis for a candidate degree in technical sciences in specialty 05.13.23 (systems and means of artificial intelligence). Kharkiv National University of Radio Electronics, Ministry of Education and Science of Ukraine, Kharkiv, 177 p.
5. Krivoyat A. G. (2020). Neural network mathematical models of sound signals in recognition tasks. - Qualification scientific work on the rights of the manuscript. Thesis for a candidate degree in physico-matematic sciences in specialty 01.05.02 "Mathematical modeling and computing methods", Zaporizhzhya National University of Ministry of Education and Science of Ukraine, Zaporozhye, 160 p.
6. Deeper, M. M., Salata, K. V., Tkach, N. A.
(2020). Construction of a diagnostic expert-medical system using neural networks, Programming problems, 2-3, 384-391.
7. Zhou, He & Huang, Jianjun & Peng, Xuemei.
(2021). Design of Medical Diagnostic System Based on Artificial Intelligence. Journal of Physics: Conference Series. 2037. 012081. 10.1088/17426596/2037/1/012081.
8. Abu-Zeid, Noha & Kashef, Rasha & Badawy, Osama. (2012). Immune Based Clustering for Medical Diagnostic Systems. Proceedings - 2012 International Conference on Advanced Computer Science Applications and Technologies, ACSAT 2012. 372-375. 10.1109/ACSAT.2012.42.
9. Xue, Qinghan & Chuah, Mooi Choo. (2019). Explainable deep learning based medical diagnostic system. Smart Health. 13. 10.1016/j.smhl.2019.03.002.
10. Sun, BQ & Zhai, AM & Feng, YJ & Wang, XF & Wang, HQ. (2003). A study on developing the intelligent medical diagnostic system. 481-483.
11. Chaturvedi, Animesh & Tiwari, Aruna & Cha-turvedi, Shubhangi & Lio, Pietro. (2022). System Neural Network: Evolution and Change based Structure Learning. IEEE Transactions on Artificial Intelligence. PP. 1-1. 10.1109/TAI.2022.3143778.
12. Li, Aiping & Jin, Songchang & Zhang, Lumin & Jia, Yan. (2015). A sequential decision-theoretic model for medical diagnostic system. Technology and Health Care. 23. S37-S42. 10.3233/thc-150926.
13. Moradpour, Amin & Karadima, Olympia & Alic, Ivan & Ragulskis, Mykolas & Kienberger, Ferry & Kosmas, Panagiotis. (2021). Development of a Solid and Flexible Matching Medium for Microwave Medical Diagnostic Systems. Diagnostics. 11. 550. 10.3390/diagnostics11030550.
14. Onisko, Agnieszka & Druzdzel, Marek. (2014). Impact of Bayesian Network Model Structure on the Accuracy of Medical Diagnostic Systems. 10.1007/978-3-319-07176-3_15.
15. Khandoker, Ahsan & Begg, Rezaul. (2008). Machine Learning for Designing an Automated Medical Diagnostic System. Handbook of Research on Modern Systems Analysis and Design Technologies and Applications. 544-559. 10.4018/978-1-59904-887-1.ch030.
16. Takata, Hideaki & Nogawa, Hiroki & Nagata, Hiroshi & Gomi, Yuichiro & Tanaka, Hiroshi. (2007). Implementation of Medical Diagnostic System Based on Epidemiological Data. Journal of Korean Society of Medical Informatics. 13. 181. 10.4258/jksmi.2007.13.2.181.
17. Akbari, Zohreh & Unland, Rainer. (2021). A Powerful Holonic and Multi-Agent-Based Front-End for Medical Diagnostics Systems. 10.1007/978-3-03079161-2 13.
