=Paper=
{{Paper
|id=Vol-2753/paper47
|storemode=property
|title=Retrospective Analysis by Multifactor Regression in the Evaluation of the Results of Fine-needle Aspiration Biopsy of Thyroid Nodules
|pdfUrl=https://ceur-ws.org/Vol-2753/short18.pdf
|volume=Vol-2753
|authors=Askold Kucher,Oksana Boyko,Kateryna Ilkanych,Andriy Fechana,Natalya Shakhovska
|dblpUrl=https://dblp.org/rec/conf/iddm/KucherBIFS20
}}
==Retrospective Analysis by Multifactor Regression in the Evaluation of the Results of Fine-needle Aspiration Biopsy of Thyroid Nodules==
https://ceur-ws.org/Vol-2753/short18.pdf
Retrospective Analysis by Multifactor Regression in the
Evaluation of the Results of Fine-needle Aspiration Biopsy of
Thyroid Nodules
Askold Kuchera, Oksana Boykoa, Kateryna Ilkanycha, Andriy Fechana,b, Natalya
Shakhovskab
a
Danylo Halytsky Lviv National Medical University, 69 Pekarska str., Lviv, 79010, Ukraine
b
Lviv Polytechnic National University, 12 Bandera str., Lviv, 79013,
Abstract
A retrospective study of the results of the examination of 160 patients who underwent fine-
needle aspiration biopsy of thyroid nodules has been conducted. Possibilities of using the TI-
RADS (Thyroid Imaging Reporting and Data System) scale as a standardized USG thyroid
examination protocol have been considered. Combinations of indicators (sex, age, ultrasound
characteristics) were analyzed with the help of multifactor regression in order to recommend
the thyroid fine-needle aspiration biopsy. Measures to increase the ultrasound database,
which will allow further use of neural networks for their analysis, have been developed.
Keywords 1
thyroid nodules, TI-RADS, fine-needle aspiration biopsy, multifactor regression
1. Introduction
Nodal thyroid formation occurs in 30-40% of the general population and is the most common
pathology among endocrine diseases. Malignant thyroid tumors account for about 3-4% of the total
number of human tumors [1]. Recently, however, due to the Chernobyl accident and radioactive
contamination of the territories, the number of thyroid tumors has increased markedly and the
tendency to increase their frequency remains [2.3].
The use of modern ultrasonic technologies allows to detect nodular formations with a size of 0.3
cm. Fine-needle aspiration biopsy (FNAB) under ultrasonographic (USG) navigation has become
widespread in the diagnosis of thyroid diseases. Cytological examination is the fastest method of
preoperative diagnosis [3].
The purpose of this study was to conduct a comparative analysis of ultrasonographic (USG)
examination of the thyroid gland with the results of cytological examination of biopsies of thyroid
nodules based on multifactorial linear regression analysis to determine indications for thyroid FNAB.
2. Materials and Methods
We have conducted a retrospective study of the results of the examination of 160 patients who
underwent FNAB of nodal thyroid formations. Data on age, sex of patients, thyroid ultrasound
protocols and cytological examination of bioptats have been analyzed. FNAB procedures of thyroid
nodules have been performed under ultrasound control (transduser10 MHz). The obtained samples
IDDM’2020: 3rd International Conference on Informatics & Data-Driven Medicine, November 19–21, 2020, Växjö, Sweden
EMAIL: kuchermd@gmail.com (A. 1); oxana_bojko@ukr.net (A. 2); ilkanych.katja@gmail.com (A. 3); andrii.v.fechan@lpnu.ua (A. 4);
nataliya.b.shakhovska@lpnu.ua (A. 5)
ORCID: 0000-0002-7511-7057 (A. 1); 0000-0002-8810-8969 (A. 2); 0000-0002-6536-7802 (A. 3); 0000-0001-9970-5497 (A. 4); 0000-
0002-6875-8534 (A. 5)
©️ 2020 Copyright for this paper by its authors.
Use permitted under Creative Commons License Attribution 4.0 International (CC BY 4.0).
CEUR Workshop Proceedings (CEUR-WS.org)
were fixed in 95% ethanol solution followed by Romanosky-Giemza staining with May-Grunwald
fixation. Cytological examination of the material has been performed by a qualified cytologist.
The ultrasonographic protocol was descriptive in accordance with the TI-RADS (Thyroid Imaging
Reporting and Data System) scale:
TI-RADS 1 –normal thyroid gland, absence of focal changes in the thyroid parenchyma;
TI-RADS 2 –formations with a low level of malignancy (benign nodules);
TI-RADS 3 – formations with an average level of malignancy (probably benign nodules);
TI-RADS 4А – formations with a high level of malignancy without changes in regional lymph
nodes;
TI-RADS 4В – formations with a high level of malignancy with changes in regional lymph nodes;
TI-RADS 5 – formations with a pre-established malignant process in it.
The cytological conclusion was descriptive in nature, which included a category from I to VI
according to the Bethesda system:
category I – nondiagnostic/ unsatisfactory bioptat;
category II – benign process;
category III – atypia of undetermined significance or follicular lesion of undetermined
significance;
category IV – follicular neoplasm or suspicious for a follicular neoplasm;
category V – suspicious for malignancy;
category VI – malignant.
Determining the relationship between ultrasonographic (USG) examination of the thyroid gland
and the results of cytological examination of biopsies of thyroid nodules was carried out by
multivariate regression analysis, which allows to make reasoned conclusions about the development
of the studied processes based on and supported by specific mathematical calculations [4].
Linear regression involves the construction of such a straight line, in which the values of the
indicators lying on it will be as close as possible to the actual, then extrapolating this line can predict
the result.
In the process of constructing multifactor regression models, the following stages can be
distinguished:
1. Selection and analysis of all possible factors that affect the process (or indicator) being
studied.
2. Mathematical and statistical analysis of factors - the main assumptions of classical regression
analysis are checked.
3. Selection the type of regression multifactor model.
4. Verifying the significance of the found parameters of the model and assessing it for the
adequacy of real reality.
The following indications have been analyzed: age, sex, shape, margin, composition, echogenicity,
conglomerate, vascularization and number of nodules.
To conclude about the feasibility of using the found model, the analysis was conducted in the
following areas:
1) The Fisher test was calculated and the found model was checked for adequacy with the source
data;
2) Variance of indicators was calculated and analyzed;
3) The correlation coefficient was calculated and analyzed;
4) The coefficient of elasticity was calculated and analyzed;
5) The confidence interval for predicted measures was calculated.
Fisher's F-statistic is calculated with m and (n-m-1) degrees of freedom:
i 1 yip yc (1)
n 2
F m
i 1 yi yip
n 2
n m 1
whеre m– the number of factors included into the model;
n – total number of observations;
уір – the calculated value of the dependent variable in the i-th observation;
ус – the average value of the dependent variable;
уі – the value of the dependent variable in the i-th observation.
According to Fisher's F-tables, the critical value Fcr with m and (n-m-1) degrees of freedom and
given confidence level is found.
The procedure for finding the regression equation between different numerical sets, usually
includes the following:
1. establishing the significance of the correlation between them;
2. the possibility of representing this dependence in the form of a mathematical expression
(regression equation).
The second stage at creation of regression model is the choose of a formula of the regression
equation. Linear multifactor regression, which describes the linear relationship between the studied
data, is written as:
y a0 a1 x1 ... an xn (2)
where у – dependent variable, function;
а0, …, аn – regression coefficients;
х1, …, хn – independent variables.
Dependent variables in linear multifactorial regression were: sex, nodule size, presence of
conglomerates, shape, contours, nodule structure, presence of calcifications, echogenicity,
vascularization, elastography, presence of lymph nodes, number of nodules.
To establish adequate coefficients, a logistic regression model with the same set of predictors was
created, but with introducing L1-regularization (least absolute shrinkage and selection operator,
LASSO) into the construction process. The essence of L1-regularization is in addition to the target
regression function a fine for the complexity of the model, proportional to the norm of the coefficient
vector. As a result, the coefficients were obtained: a0= 0,3699845; a1= 0,4097208; a2= -0,135977;
a3= -0,5495012; a4= -0,0949706; a5= 0,154647; a6= 0,0502302; a7= 0,6084696; a8= -0,2844752;
a9= 0,4817886; a10= 0,0574823; a11= 0,4017387; a12= 1,3212888.
3. Results
The results of ultrasonography (USG).
The results of USG in 160 patients have been analyzed. The material from 226 thyroid nodules
was obtained by FNAB method. According to USG, 4 nodules are assigned to the TI-RADS 2
category; 145 nodules to the TI-RADS 3 category; 77 nodules to category TI-RADS 4.
The results of FNAB.
According to the results of cytological examination of 226 biopsies, 5 samples were assigned to
category I, 119 to category II, 44 to category III; 24 to category IV, 15 to category V, 19 to category
VI.
Comparison of USG results and cytological examination of thyroid nodules is presented in
Table.1.
4. Discussion
Among 145 nodules that were assigned to TI-RADS 3 (formations with an average level of
malignancy - probably benign nodules) in accordance with the TI-RADS scale, the benign process
(Bethesda II) was confirmed in 88 (60.7%) cases. Given the large percentage (60.7%) of benign
nodules that were suspected of being malignant on the TI-RADS scale, it can be assumed that there is
a certain combination of indicators (sex, age, ultrasound characteristics) that will allow monitoring of
the nodule and abandoning FNAB in first detected nodule.
Table 1
Comparison of USG results and cytological examination of thyroid nodules
Cytological TI-RADS 2 TI-RADS 3 TI-RADS 4
conclusion (summarily 4А
and 4В)
Bethesda I 0 4 (2,7%) 1(1,3%)
(unsatisfactory bioptat)
Bethesda II (benign process) 3(75,0%) 88 (60,7%) 28 (36,4%)
Bethesda III (atypia of 1 (25,0%) 29 (20,0%) 14 (18,2%)
undetermined significance or
follicular lesion of
undetermined significance)
Bethesda IV (follicular 0 13 (8,9%) 11 (14,3%)
neoplasm or suspicious for a
follicular neoplasm)
Bethesda V 0 9 (6,2%) 6 (7,8%)
(suspicious for malignancy)
Bethesda VI (malignant) 0 2 (1,4%) 17 (22,1%)
Total 4 (100%) 145 (100%) 77 (100%)
Unequivocally, it can be argued that it is very important to find certain combinations, which is
typical for nodules that correspond to the TI-RADS 3, but were classified as Bethesda III, IV, V and
VI.
Among the 77 nodules that were assigned to TI-RADS 4 (formations with a high level of
malignancy) on the TI-RADS scale, the benign process was confirmed in 28 (36.4%) cases. This once
again confirms the importance of finding a certain combination of indicators in nodules corresponding
to the cytological Bethesda II. In turn, if the nodule corresponds to the TI-RADS 4 in according with
the USG, it is mandatory to perform FNAB, and not immediately perform surgery.
5. Neural Network Applications Perspectives
Nowadays, neural networks are a promising area of research in medicine, especially in the field of
oncology diagnostics [5-6]. Consider the prospects for the application of neural networks in the
medical field of oncology diagnosis in a preclinical study. In carcinogenesis, artificial neural networks
are successfully applied to the problems of both preclinical and post-clinical diagnoses. The primary
goal of medical diagnostic research is to develop more cost-effective and easy-to-use systems,
procedures, and methods to support clinicians. In our case, neural networks can be used to analyze
ultrasound results in order to develop diagnostic algorithms that can improve data sorting practices.
The neural network plays an important role in the decision support system. Preliminary data
analysis suggests that the optimal system for our analysis would be a 5-layer model with 6 outputs
and 37 inputs. However, nowadays, there is a critical lack of data for this analysis. To solve this
problem, we propose to create a database of ultrasound results with remote access.
The specifics of its work is that third-party users after a preliminary verification will be able to
enter their own data into the general database. To facilitate our work, we use a browser application
that easily adapts to work with stationary and mobile devices. Figer 1 shows the interface view of the
first user login.
The process of filling in the results of studies looks like a consistent selection of options from the
drop-down windows. The first entry involves answering several narrowly professional questions to
avoid clogging registration by unauthorized persons. In the future, the user is provided with a personal
password and a code that will accompany his data in the database.
Figure 1: The interface of the first input to fill the ultrasound database.
6. Conclusions
• The use of the TI-RADS scale is recommended for mandatory use in order to standardize the
USG thyroid examination protocol.
• Cytological examination of biopsies of thyroid nodules with assessment by the Bethesda system
is an informative method for diagnosing cancer.
• In order to determine the indications for thyroid FNAB, it is recommended to continue searching
for combinations of criteria that could clearly recommend this manipulation.
• Measures to increase the ultrasound database, which will allow further use of neural networks for
their analysis have been developed.
[1] Cristina Romei, Raffaele Ciampi and Rossella Elisei “A comprehensive overview of the role of
the RET proto-oncogene in thyroid carcinoma”. Nature Reviews Endocrinology, 12.4, (2016):
192-202.
[2] M. Tronko, T. Bogdanova, L. Voskoboynyk, L. Zurnadzhy, V. Shpak and L. Gulak. “Radiation
induced thyroid cancer: fundamental and applied aspects”. Experimental oncology. 32/3 (2010):
200-204.
[3] Gabriella Pellegriti, Francesco Frasca, Concetto Regalbuto, Sebastiano Squatrito and Riccardo
Vigneri. “Worldwide increasing incidence of thyroid cancer: update on epidemiology and risk
factors”. Journal of cancer epidemiology, 2013 (2013): 965212
https://doi.org/10.1155/2013/965212.
[4] Guiwen Rong, Xu Wang, Hui Xu and Baiqing Xiao. Multifactor Regression Analysis for
Predicting Embankment Dam Breaching Parameters. Journal of Hydraulic Engineering, 146(2)
(2020): 04019051
[5] Athanasios V. Vasilakos, Yu Tang and Yuanzhe Yao. “Neural networks for computer-aided
diagnosis in medicine: A review”. Neurocomputing. 216. (2016): 700-708.
https://doi.org/10.1016/j.neucom.2016.08.039
[6] Benjamin H. Kann, Reid Thompson, Charles R. Thomas, Jr, Adam Dicker and Sanjay Aneja.
Artificial intelligence in oncology: current applications and future directions. Oncology, 33(2).
(2019).