=Paper=
{{Paper
|id=Vol-3609/paper15
|storemode=property
|title=Identification of Personality Based on the Sphenoid Sinus Structure Using Machine Learning
|pdfUrl=https://ceur-ws.org/Vol-3609/short2.pdf
|volume=Vol-3609
|authors=Alina Nechyporenko,Marcus Frohme,Vladyslav Omelchenko,Victoriia Alekseeva,Andrii Lupyr,Vitaliy Gargin
|dblpUrl=https://dblp.org/rec/conf/iddm/NechyporenkoFOA23
}}
==Identification of Personality Based on the Sphenoid Sinus Structure Using Machine Learning==
https://ceur-ws.org/Vol-3609/short2.pdf
Identification of Personality Based on the Sphenoid Sinus
Structure Using Machine Learning
Alina Nechyporenkoa,b, Marcus Frohmeb , Vladyslav Omelchenkob, Victoriia Alekseevab,c,d,
Andrii Lupyrc and Vitaliy Garginc,d
a
Kharkiv National University of Radioelectronics, Nauky avenue 14, Kharkiv, 61166, Ukraine
b
Technical University of Applied Sciences Wildau (TH Wildau), Hochschulring 1, Wildau, 15745, Germany
c
Kharkiv National Medical University, Nauky avenue 4, Kharkiv, 61022, Ukraine
d
Kharkiv International Medical University, Molochna street 38, Kharkiv, 61001, Ukraine
Abstract
The aim of our study is to develop a new, simple, and effective method for identification of
personality based on the characteristics of the sphenoid sinus structure, using machine
learning for subsequent implementation into routine medical practice in Ukraine. The study
involved 200 multislice computed tomography (MSCT) scans of individuals of various
genders and ages. During the study, we obtained results with an accuracy exceeding 70%.
Keywords 1
Identification of personality, multislice computed tomography, deep learning
1. Introduction
The full-scale Russian invasion has had a profound impact on Ukrainian society, presenting
numerous challenges and causing immense suffering for millions of Ukrainians [1]. War crimes
continue to be committed in the occupied territories, and the discovery of unmarked mass graves in
various locations is deeply disturbing [2]. Establishing the identities of those who have fallen victim
to Russian aggression is a crucial task at hand [3]. Personal identification is particularly important
during times of war, although no method can guarantee a 100% reliable result [1]. Fingerprints [4] are
commonly used, but in wartime, bodies may be burned or damaged, rendering fingerprint
identification impossible. Autolysis, the natural decomposition of bodies over time, can also hinder
the use of fingerprints or retinas for identification. DNA identification [5] is a promising and accurate
method, but obtaining DNA samples from deceased individuals' relatives is not always feasible.
While global databases exist for DNA identification, Ukraine lacks such a resource. Additionally,
DNA collection requires significant time, effort, and invasiveness, making implementation
challenging. The proposed study aims to utilize existing data for collection and analysis. Bones are
considered the most stable structures for the analysis [6], and studying cranial bones, particularly the
sphenoid sinus, shows promise. Computed tomography [7] (CT) scans can be used to examine the
sphenoid sinus, as it is less likely to be damaged due to its deep location within the skull. Medical
image segmentation [8], a process that identifies pixels of interest in medical images, can be
employed to process CT images. Convolutional neural networks (CNNs), particularly the U-Net
architecture, have proven effective for medical image segmentation. CNNs offer promising potential
for automated diagnostic methods and personal identification with help of deep learning [9]. While
there are existing segmentation platforms, a clear workflow and necessary functions for easy
IDDM’2023: 6th International Conference on Informatics & Data-Driven Medicine, November 17 - 19, 2023, Bratislava, Slovakia
EMAIL: alinanechiporenko@gmail.com; mfrohme@th-wildau.de; vladyslav.omelchenko3@nure.ua; vik1305230@gmail.com;
lupyr_ent@ukr.net; vitgarg@ukr.net
ORCID: 0000-0001-9063-2682; 0000-0002-4501-7426; 0009-00057395-3982, 0000-0001-5272-8704; 0000–0002–9465–224X; 0000-0001-
8194-4019,
© 2023 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)
CEUR
ceur-ws.org
Workshop ISSN 1613-0073
Proceedings
configuration are currently lacking. Therefore, CNNs present a promising approach for processing
medical CT images in the context of personal identification [10].
To date, there is a relatively small amount of scientific works dedicated to the identification of
personality. This can be attributed to several factors, including the complexity of developing
algorithms for personality‘s identification, the intricate process of pre-processing medical images, and
the limited size of databases used for solving this task. All of these factors can create additional
barriers to the implementation of methods described by various authors into medical practice, and it
appears to be a reason why these methods have not been realized.
Although the proposed method offers numerous advantages associated with accurate personal
identification, there are certain aspects that may pose challenges in achieving the intended goal. The
first drawback could be the lack of results from CT scans. Despite the availability of extensive X-ray
data covering a significant portion of Ukraine's population, approximately 30% of individuals have
not undergone CT scans according to various sources. Consequently, alternative methods should be
considered for identifying this group of people.
The second drawback could be the age factor, specifically concerning children. In our proposed
method, similar to other studies, children were not included in the groups of individuals whose
identity was intended to be identified. Considering the age-related variability of the maxillary sinus
and the infrequent use of CT scans in children due to limited indications, personal identification in
pediatric age groups can be challenging.
The aim of our study is to develop a new, simple, and effective method for identification of
personality based on the characteristics of the sphenoid sinus structure, using deep learning for
subsequent implementation into routine medical practice in Ukraine.
2. Material and Methods
The study involved analyzing CT scans of 200 individuals of various ages and genders.
Segmentation, which is the process of identifying and analyzing specific structures within medical
images, is a challenging task in medical image analysis. It provides valuable information about the
shape and characteristics of organs and tissues. To assist with the segmentation process, a Standard
Operating Procedure (SOP) was developed specifically for creating 3D models of the sphenoidal
sinus from CT images.
The procedure consists of three main steps. Firstly, the CT images are prepared, converted, and
masks are created. Secondly, models for the experiment are constructed using the popular Tensorflow
library. Finally, the experiment is conducted, and the results are compared. The RadiAnt software will
be used to select tomographic images from a Toshiba Aquilion CT [11] scanner, which can collect
data from four slices simultaneously with a high level of performance. This scanner provides high-
resolution multi-slice scanning with wide bandwidth.
The data used in this study will be obtained from 2000 individuals who have undergone CT. The
data was be used in accordance with the existing contract between Kharkiv Institute of Emergency
Medicine and the Kharkiv National Medical University. Each scan contains patient identity data, such
as name, age, sex, date of birth, occupation, referring physician, date of examination, and unique
DICOM identifiers [12].
These methods require large and accurately labeled training datasets, which are created by medical
experts. Therefore, it is crucial to choose an appropriate image annotation tool. The Labelme tool was
selected for its high accuracy and ability to output annotations in JSON format [13].
All individuals whose CT images are included in this research have voluntarily given informed
consent to participate.
The network was trained end-to-end using binary cross-entropy loss function, Adam optimizer,
learning rate 0.01, and Exponential Linear Unit (ELU) as an activation function. The size of input
image is 128x128x3 and its mask is 128x128. Epochs, batches sizes and validation split were
customized utilizing remote server with the following configuration CPU: 2x AMD EPYC 7413 24-
Core CPU, 180W, 2.65GHz, 128MB, - L3 Cache, DDR4-3200, Turbo Core max. 3.60GHz, GPU: 4x
NVIDIA Tesla A100 (NVLink), 80GB, RAM: DDR4-3200512GB.
Figure 1: MSCT slice of the Sphenoid sinus and its mask
3. Results
All the results of the experiment using U-Net model with different parameters: batch size, number
of epochs, validation split and with IoU (Intersection over Union) metric value are in the Table 1.
Table 1
Results of the experiment
Model name Batch size Number of Validation IoU
epochs split
modelA_5_10_10 5 10 0.10 0.637
modelA_5_10_33 5 10 0.33 0.617
modelA_5_15_10 5 15 0.10 0.737
modelA_5_15_33 5 15 0.33 0.658
modelB_10_10_10 10 10 0.10 0.512
modelB_10_10_33 10 10 0.33 0.441
modelB_10_15_10 10 15 0.10 0.653
modelB_10_15_33 10 15 0.33 0.705
modelC_20_10_10 20 10 0.10 0.613
The U-Net model is composed of convolutional and transposed convolutional layers. It has the
following structure: The input data, representing an image, is passed through an Input layer. The pixel
values of the image are normalized using a Lambda layer, where each pixel value is divided by 255 to
obtain normalized values ranging from 0 to 1.
Next, the input data is processed through a Conv2D layer with 32 filters. This layer utilizes the
ELU (Exponential Linear Unit) activation function, which takes into account negative values when
activating neurons.
The layer weights are initialized using the "he_normal" method, promoting more effective model
training. To prevent overfitting, a Dropout layer is applied, randomly disabling a certain percentage of
neurons (in this case, 10% of neurons with a dropout rate of 0.1). This helps reduce correlation among
neurons and enhance the model's generalization ability. Following the Dropout layer, another Conv2D
layer with 32 filters and the ELU activation function is applied. This additional convolutional layer
aids in extracting higher-level features from the image. Subsequently, a MaxPooling2D layer is used
to perform pooling operations, reducing the data dimensionality by removing redundant information
and focusing on the most important image features. Similar operations are repeated for subsequent
levels (c2, c3, c4, c5), where additional convolutional layers with an increased number of filters are
applied.
Figure 2: An example of the personality identification using various models
This allows the model to extract increasingly abstract and complex features from the image. After the
final convolutional layer, c5, the decoding process begins. To increase the data size, a
Conv2DTranspose layer is utilized, performing the reverse operation of convolution and expanding
the spatial dimensions of the data. The output of the Conv2DTranspose layer is then concatenated
with the corresponding layer (c4) using the concatenate layer. This allows for preserving spatial
dependencies between features and connecting them for further decoding. Additional convolutional
layers, c6, are then applied for decoding and feature extraction at different levels. The decoding
operations are repeated for subsequent levels (c7, c8, c9), gradually reducing the number of filters in
the convolutional layers. This enables the model to progressively restore the original data size and
refine the results. Finally, a Conv2D layer with a single filter and the sigmoid activation function is
used. This layer is employed to obtain the final segmented image, where each pixel takes a value from
0 to 1, representing the probability of belonging to a specific class. The model is compiled using the
Adam optimizer and binary_crossentropy loss function. The Adam optimizer efficiently updates the
model weights by minimizing the loss function. The binary_crossentropy loss function is employed
for binary classification tasks, where the goal is to separate image pixels into two classes: background
and object.
4. Discussion
Thus, we have developed a new approach to personal identity identification using machine
learning. It is worth noting that the proposed method differs from similar approaches and offers
several advantages.
Currently, one of the most promising research studies is the work of Wen H. et al. [14] and Dong
et al. [15], The studies were conducted on a large database of subjects (732 individuals). However, it
is worth noting the relatively uneven distribution between the training group and the target group (600
and 123 individuals, respectively). Furthermore, in the course of this research, all patients were
categorized into ranks (from 1 to 5) based on the "complexity" of their sphenoid sinus structures. Our
own experiment demonstrates that it is most challenging to identify sinuses with smaller volumes and
simple configurations (lacking additional septa, bays, and cellular structure). In our opinion, such
categorization could potentially distort the results.
Another negative factor in both of the presented works is the requirement for extensive pre-
processing in the identification process. Given the substantial number of individuals whose identities
need to be verified in Ukraine and the shortage of medical and technical personnel capable of serving
as experts for this task, the process of implementing a personal identity identification protocol into the
country's healthcare system could consume a significant amount of time.
In the study presented by Soudhin et al [16], it is notable that a small number of subjects were
included in the experiment (72 individuals). Thirteen of these subjects underwent CT scans twice. The
inclusion of such individuals in both the target and training groups, given the relatively small number
of subjects in each group, can intentionally increase the percentage of positive results.
A common drawback of other studies known to date is the limited number of subjects in the
database [17, 18].
During the experiment conducted on a large dataset of 200 individuals, results were obtained that
can aid in the identification of a person, whether alive or deceased. The proposed method is
straightforward, and its implementation only requires the assessment of a medical expert to evaluate
the quality of image labeling in the training group. This method is versatile, as it can be easily
integrated into the healthcare system of any country, not limited to Ukraine. The method holds
promise, as it has significant potential for improving the accuracy of identity determination. For
instance, while it has been developed for 2D CT scans for now, there are plans to extend the analysis
to 3D images in the future. Furthermore, the study aims to include anatomical structures (such as the
optic nerve canal, internal carotid artery canal, and optic nerve canal) known for their notable
variability, which has not been previously included in any similar studies [19] or even studies in an
other medical fields [20, 21]. Using 3D segmentation is more rational, considering the fact that three-
dimensional imaging most accurately captures the anatomical features of the area under investigation
[22]. It is known that automatic and semi-automatic methods can be used for image annotation. Most
authors prefer automatic research [23] methods without involving medical professionals in the
annotation process. However, given the complexity and diversity of the sinus structure, physician
oversight may be necessary and could enhance the effectiveness of the study.
5. Conclusions
In conclusion, our study has led to the development of a straightforward, informative, and readily
applicable method for personal identity identification. This method can be seamlessly integrated into
medical practice and utilized for identifying individuals, both living and deceased, not only in Ukraine
but also in any country around the world. Furthermore, the inclusion of physician oversight in the
annotation process adds an extra layer of reliability to our methodology.
The implications of our findings are significant, offering potential applications in various fields
such as forensics, anthropology, and medical research. The simplicity and efficiency of our method
make it accessible for implementation in diverse healthcare settings, promising a practical solution for
personal identification challenges. As our research continues to advance, we anticipate further
refinements and applications, ultimately contributing to the enhancement of identification techniques
on a global scale
.
6. References
[1] Pavlova, I., Graf-Vlachy, L., Petrytsa, P., Wang, S., & Zhang, S. X. (2022). Early evidence on
the mental health of Ukrainian civilian and professional combatants during the Russian invasion.
European psychiatry : the journal of the Association of European Psychiatrists, 65(1), e79.
https://doi.org/10.1192/j.eurpsy.2022.2335
[2] Ashbridge, S. I., Randolph-Quinney, P. S., Janaway, R. C., Forbes, S. L., & Ivshina, O. (2022).
"Environmental conditions and bodily decomposition: Implications for long-term management of
war fatalities and the identification of the dead during the ongoing Ukrainian conflict." Forensic
Science International: Synergy, 5, 100284. [Online]. Available:
https://doi.org/10.1016/j.fsisyn.2022.100284.
[3] Hăisan, A., Măirean, C., Lupuşoru, S. I., Tărniceriu, C., & Cimpoeşu, D. (2022). "General Health
among Eastern Romanian Emergency Medicine Personnel during the Russian-Ukrainian Armed
Conflict." Healthcare (Basel, Switzerland), 10(10), 1976. [Online]. Available:
https://doi.org/10.3390/healthcare10101976.
[4] Champod, C. (2015). "Fingerprint Identification: Advances since the 2009 National Research
Council Report." Philosophical Transactions of the Royal Society of London. Series B,
Biological Sciences, 370(1674), 20140259. [Online]. Available:
https://doi.org/10.1098/rstb.2014.0259.
[5] Madsen, D., Azevedo, C., Micco, I., Petersen, L. K., & Hansen, N. J. V. (2020). "An Overview
of DNA-Encoded Libraries: A Versatile Tool for Drug Discovery." Progress in Medicinal
Chemistry, 59, 181–249. [Online]. Available: https://doi.org/10.1016/bs.pmch.2020.03.001.
[6] A. Nechyporenko, V. Alekseeva, R. Nazaryan, and V. Gargin, “Biometric Recognition of
Personality based on Spiral Computed Tomography Data,” 2021 IEEE 16th International
Conference on the Experience of Designing and Application of CAD Systems (CADSM). IEEE,
Feb. 22, 2021. doi: 10.1109/cadsm52681.2021.9385267.
[7] Brough, A. L., Morgan, B., & Rutty, G. N. (2015). "Postmortem Computed Tomography
(PMCT) and Disaster Victim Identification." La Radiologia Medica, 120(9), 866–873. [Online].
Available: https://doi.org/10.1007/s11547-015-0556-7.
[8] V. Alekseeva, A. Nechyporenko, M. Frohme, V. Gargin, I. Meniailov, and D. Chumachenko,
“Intelligent Decision Support System for Differential Diagnosis of Chronic Odontogenic
Rhinosinusitis Based on U-Net Segmentation,” Electronics, vol. 12, no. 5. MDPI AG, p. 1202,
Mar. 02, 2023. doi: 10.3390/electronics12051202.
[9] S. S. Mondal, N. Mandal, K. K. Singh, A. Singh, and I. Izonin, “EDLDR: An Ensemble Deep
Learning Technique for Detection and Classification of Diabetic Retinopathy,” Diagnostics, vol.
13, no. 1. MDPI AG, p. 124, Dec. 30, 2022. doi: 10.3390/diagnostics13010124.
[10] Igarashi, Y., Kondo, S., Kida, S., Aibara, M., Kaneko, M., & Uchikoba, F. (2022). "Mandibular
Premolar Identification System Based on a Deep Learning Model." Journal of Oral Biosciences,
64(3), 321–328. [Online]. Available: https://doi.org/10.1016/j.job.2022.05.005.
[11] . Freislederer, P., Heinz, C., von Zimmermann, H., Gerum, S., Roeder, F., Reiner, M., Söhn, M.,
Belka, C., & Parodi, K. (2018). "Clinical Workflow Optimization to Improve 4DCT
Reconstruction for Toshiba Aquilion CT Scanners." Zeitschrift fur medizinische Physik, 28(2),
88–95. [Online]. Available: https://doi.org/10.1016/j.zemedi.2017.12.003.
[12] Brühschwein, A., Klever, J., Hoffmann, A. S., Huber, D., Kaufmann, E., Reese, S., & Meyer-
Lindenberg, A. (2020). "Free DICOM-Viewers for Veterinary Medicine: Survey and
Comparison of Functionality and User-Friendliness of Medical Imaging PACS-DICOM-Viewer
Freeware for Specific Use in Veterinary Medicine Practices." Journal of Digital Imaging, 33(1),
54–63. [Online]. Available: https://doi.org/10.1007/s10278-019-00194-3.
[13] Liu, J., Yang, M., Zhang, L., & Zhou, W. (2019). "An Effective Biomedical Data Migration Tool
from Resource Description Framework to JSON." Database: The Journal of Biological Databases
and Curation, 2019, baz088. [Online]. Available: https://doi.org/10.1093/database/baz088.
[14] Wen, H., Wu, W., Fan, F., Liao, P., Chen, H., Zhang, Y., Deng, Z., & Lv, W. (2022). "Human
Identification Performed with Skull's Sphenoid Sinus Based on Deep Learning." International
Journal of Legal Medicine, 136(4), 1067–1074. https://doi.org/10.1007/s00414-021-02761-2.
[15] Dong X, Fan F, Wu W, Wen H, Chen H, Zhang K, Zhang J, Deng Z. (2022). "Forensic
Identification from Three-Dimensional Sphenoid Sinus Images Using the Iterative Closest Point
Algorithm." Journal of Digital Imaging, 35(4), 1034-1040. https://doi.org/10.1007/s10278-021-
00572-w.
[16] Souadih, K., Belaid, A., Ben Salem, D., & Conze, P. H. (2020). "Automatic forensic
identification using 3D sphenoid sinus segmentation and deep characterization." Medical &
Biological Engineering & Computing, 58(2), 291–306. https://doi.org/10.1007/s11517-019-
02050-6.
[17] Deloire, L., Diallo, I., Cadieu, R., Auffret, M., Alavi, Z., Ognard, J., & Ben Salem, D. (2019).
"Post-mortem X-ray computed tomography (PMCT) identification using ante-mortem CT-scan
of the sphenoid sinus." Journal of Neuroradiology = Journal de Neuroradiologie, 46(4), 248–255.
https://doi.org/10.1016/j.neurad.2018.08.003.
[18] Auffret, M., Garetier, M., Diallo, I., Aho, S., & Ben Salem, D. (2016). "Contribution of
computed tomography to the anatomical aspects of the sphenoid sinuses for forensic
identification." Journal of Neuroradiology = Journal de Neuroradiologie, 43(6), 404–414.
https://doi.org/10.1016/j.neurad.2016.03.007.
[19] D, P., Prabhu, L. V., Kumar, A., Pai, M. M., & Kvn, D. (2015). "The anatomical variations in the
neurovascular relations of the sphenoid sinus: an evaluation by coronal computed tomography."
Turkish Neurosurgery, 25(2), 289–293. https://doi.org/10.5137/1019-5149.JTN.10638-14.
[20] Chumachenko D., Meniailov I., Bazilevych K., Chumachenko T., Yakovlev S. Investigation of
Statistical Machine Learning Models for COVID-19 Epidemic Process Simulation: Random
Forest, K-Nearest Neighbors, Gradient Boosting, Computation, 2022, vol. 10, iss. 6, 86. doi:
10.3390/computation10060086
[21] Chumachenko D. On intelligent multiagent approach to viral Hepatitis epidemic processes
simulation, Proceedings of the 2018 IEEE 2nd International Conference on Data Stream Mining
and Processing, DSMP 2018, 2018, pp. 415-419. doi: 10.1109/DSMP.2018.8478602
[22] K. Souadih, A. Belaid, D. Ben Salem, and P.-H. Conze, “Automatic forensic identification using
3D sphenoid sinus segmentation and deep characterization,” Medical & Biological
Engineering & Computing, vol. 58, no. 2. Springer Science and Business Media LLC, pp.
291–306, Dec. 17, 2019. doi: 10.1007/s11517-019-02050-6
[23] P. Maken, A. Gupta, and M. K. Gupta, “A systematic review of the techniques for automatic
segmentation of the human upper airway using volumetric images,” Medical & Biological
Engineering & Computing, vol. 61, no. 8. Springer Science and Business Media LLC, pp.
1901–1927, May 30, 2023. doi: 10.1007/s11517-023-02842-x.