|
| |
 |
|
| ORIGINAL ARTICLE |
|
| Year : 2022 | Volume
: 71
| Issue : 3 | Page : 204-209 |
|
Gender prediction with the parameters obtained from pelvis computed tomography images and machine learning algorithms
Yusuf Secgin1, Zulal Oner2, Muhammed Kamil Turan3, Serkan Oner4
1 Department of Anatomy, Faculty of Medicine, Karabük University, Karabük, Turkey
2 Department of Anatomy, Faculty of Medicine, İzmir Bakırçay University, İzmir, Turkey
3 Department of Medical Biology, Faculty of Medicine, Karabük University, Karabük, Turkey
4 Department of Radiology, Faculty of Medicine, İzmir Bakırçay University, İzmir, Turkey
| Date of Submission | 30-Dec-2020 |
| Date of Decision | 06-Jun-2022 |
| Date of Acceptance | 09-Jun-2022 |
| Date of Web Publication | 20-Sep-2022 |
Correspondence Address:
Dr. Zulal Oner
Department of Anatomy, Faculty of Medicine, İzmir Bakırçay University, İzmir
Turkey
 Source of Support: None, Conflict of Interest: None
DOI: 10.4103/jasi.jasi_280_20
Introduction: In the skeletal system, the most dimorphic bones employed for postmortem gender prediction include the bones in the pelvic skeleton. Bone measurements are usually conducted with cadaver bones. Computed tomography (CT) is an increasingly popular method due to its ease of use, reconstruction opportunities, and lower impact of age bias and provides a modern data source. Even when parameters obtained with different or same bones are missing, machine learning (ML) algorithms allow the use of statistical methods to predict gender. This study was carried out in order to obtain high accuracy in estimating gender with the pelvis skeleton by integrating ML algorithms, which are used extensively in the field of engineering, in the field of health. Material and Methods: In the present study, pelvic CT images of 300 healthy individuals (150 females, 150 males) between the ages of 25 and 50 (the mean female age = 40, the mean male age = 37) were transformed into orthogonal images, and landmarks were placed on promontory, iliac crest, sacroiliac joint, anterior superior iliac spine, anterior inferior iliac spine, terminal line, obturator foramen, greater trochanter, lesser trochanter, femoral head, femoral neck, body of femur, ischial tuberosity, acetabulum, and pubic symphysis, and coordinates of these regions were obtained. Four groups were formed based on various angle and length combinations obtained from these coordinates. These four groups were analyzed with ML algorithms such as Logistic Regression, Linear Discriminant Analysis (LDA), Random Forest, Extra Trees Classifier, and ADA Boost Classifier. Results: In the analysis, it was determined that the highest accuracy was 0.96 (sensitivity 0.95, specificity 0.97, Matthew's Correlation Coefficient 0.93) with LDA. Discussion and Conclusion: The use of length and angle measurements obtained from the pelvis showed that the LDA model was effective in estimating gender.
Keywords: Computed tomography, gender prediction, machine learning algorithms, pelvis
How to cite this article:
Secgin Y, Oner Z, Turan MK, Oner S. Gender prediction with the parameters obtained from pelvis computed tomography images and machine learning algorithms. J Anat Soc India 2022;71:204-9 |
How to cite this URL:
Secgin Y, Oner Z, Turan MK, Oner S. Gender prediction with the parameters obtained from pelvis computed tomography images and machine learning algorithms. J Anat Soc India [serial online] 2022 [cited 2023 Mar 28];71:204-9. Available from: https://www.jasi.org.in/text.asp?2022/71/3/204/356496 |
| Introduction | |  |
Gender prediction is the foundation of forensic science and anthropology. Gender prediction is the first step in identity determination, including factors such as age, height, weight, etc.[1] Gender could be predicted with osteometry, odontometrics, and DNA analysis. DNA analysis is expensive and not commonly available and requires qualified staff. Furthermore, osteometry is a reliable, preferred, inexpensive, and highly accessible method that provides fast results, and does not require qualified staff.[2]
Almost all parts of the human skeleton have been studied for gender prediction, and it was observed that the most dimorphic region was the pelvic skeleton and cranium.[3],[4] Certain studies reported that the pelvic skeleton was more dimorphic than the cranium.[5] One of the key factors that led to the determination of the pelvis skeleton as the most dimorphic region was the impact of sex hormones (estrogen and androgen) on the human skeleton. Sex hormones affect the pelvic skeleton the most. Because estrogen leads a female pelvis suitable for birth, androgen leads to one that could support the muscle mass and tonus in men.[3]
Multidetector computed tomography (MDCT) is a modern imaging method that could distinct bone tissue and other tissues. Thin section images lead to the determination of 3-dimensional image direction and the image could be positioned vertically. This feature allows the determination of a more realistic coordinate, angle, and length.[2] Computed tomography (CT) is preferred in postmortem cases with soft tissue loss or partial bone deformities since it allows easy, fast, and low-cost reconstruction.[6]
There are three machine learning (ML) algorithm types. The first is the controlled type that models the correlations between the inputs and outputs, the second lass is the uncontrolled type that can provide outputs based on previously unknown data, and the third is the amplified type that matches the inputs with the desired outputs.[7] The present study aimed to obtain accurate results based on various pelvic skeleton parameters with ML algorithms that offer new horizons in the field of medicine.
| Material and Methods | | |
This section includes details on image sample sources, image transformation to the orthogonal plane, the collection of the parameters, and employed ML algorithms.
Image samples
The present study was approved by the ethics committee decision no: 6/23. In the present retrospective study, pelvic CT images of 300 randomly selected individuals (150 females, 150 males), who were admitted to the hospital with various indications and without pelvic skeleton fracture or any related pathology were employed. The mean female participant age was 40 and the mean male participant age was 37. The Anderson-Darling normality test was used to determine whether the participant age exhibited normal distribution, and it was determined that the participant age did not exhibit a normal distribution. A significant difference was found between participant age with the Mann–Whitney U-test (P ≤ 0.001).
Multidetector computed tomography protocol
The supine pelvic CT images were taken with a 16-row MDCT scanner (Aquilion 16; Toshiba Medical Systems, Otawara, Japan) with a section thickness of 5 mm. Scanning protocol values were as follows: tube voltage: 120 kV, gantry rotation: 0.75 s, and pitch: 1.0 mm.
Image Analysis
The pelvic CT images were recorded on Horos Medical Image Viewer 3.0 (United States) software in Digital Imaging and Communications in Medicine format. Three dimensional Curved Multiplanar Reconstruction was conducted on the recorded images to obtain 3-planar images (coronal, sagittal, transversal). All images were transformed into an orthogonal plane by matching the coronal, sagittal, and transversal image series and the promontory plane. All sections transformed into the orthogonal plane were superimposed with Horos software to obtain a single image [Figure 1]. | Figure 1: (a) The transversal image in orthogonal form, (b) The sagittal image in orthogonal form, (c) Coronal image in the orthogonal form (d) Superposed image
Click here to view |
Landmarks were placed on orthogonal pelvic CT image promontory, iliac crest, sacroiliac joint, anterior superior iliac spine, anterior inferior iliac spine, terminal line, obturator foramen (for), greater trochanter, lesser trochanter, femoral head, femoral neck, body of femur, ischial tuberosity, acetabulum, and pubic symphysis to obtain the coordinates for these anatomic locations [Figure 2]. | Figure 2: (a) promontory, (b) iliac crest, (c) sacroiliac joint, (d) anterior superior iliac spine, (e) anterior inferior iliac spine, (f) sacroiliac joint lower end, (g) terminal line, (h) acetabulum upper end, (i) acetabulum lower end, (k) greater trochanter, (n) femoral head, (o) femoral neck, (p) body of femur, (r) obturator foramen upper end, (s) obturator foramen lower end, (u) ischial tuberosity, (t) pubic symphysis, (m) lesser trochanter
Click here to view |
These coordinates were transferred to a graphics software developed by Karabük University faculty members in Medical Biology, Anatomy, Radiology departments called Sekazu (Version 3.0, Karabük, Turkey) which could calculate mathematical concepts such as length, angle, and area automatically and could analyze gender prediction with ML algorithm.
Four groups were formed based on the mathematical units such as length and angle and the transferred coordinates. The determining lengths and angles are presented in [Table 1].
Various ML algorithms (Linear Discriminant Analysis [LDA], Random Forest [RF], Logistic Regression [LR], Extra Trees Classifier [ETC], ADA Boost Classifier [ADA]) were applied to the above-mentioned mathematical elements and the groups with an accuracy (ACC) of 90% or over were selected. The ACC of the groups that were not selected was between 80 and 90%. The 4 groups created based on the determined angles and lengths are presented in [Table 2].
Machine learning algorithms
The RF algorithm, which is fast and load resistant, was introduced by Brierman in 2001 and is a community learning algorithm that collects the decisions of several independent multivariate trees.[8]
The LR algorithm is employed to reveal the unique effect of one or more independent variables on the result. Due to its ability to predict the outcome and to reveal the unique effect of each data on the result, it is preferred in medical research.[9]
The LDA is an analysis method commonly used by anthropologists, which could discriminate various classes, is easy to access and implement, and could categorize new classes based on the input data.[10],[11]
Although the ETC algorithm is similar to RF, it is a proven method when compared to RF due to stronger variance. There are two important differences between ETC and RF; it employs all data in the training series, and the nodes are randomly divided.[12]
ADA is a high-result method that provides strong classifier algorithms by integrating weak classifier algorithms that are proven to yield stronger results.[13]
The results were obtained by applying ML algorithms to the 4 groups.
Performance criteria
These included accuracy (Acc), sensitivity (Sen), specificity (Spe), F1, and Matthews Correlation Coefficient (Mcc) obtained with the confusion matrix.
Equation 1. (TP: True positive, TN: True negative, FP: False positive, FN; False negative).
| Results | | |
In the present study, 5 ML algorithms were applied separately to 4 groups, and Acc, Spe, Sen, Mcc were obtained [Table 3], [Table 4], [Table 5], [Table 6]. The highest Acc was 0.96 determined with the LDA algorithm applied to the first group. The mean Acc obtained with 5 ML algorithms applied to the four groups was 0.91. The highest Sen was 0.96 with the ADA algorithm applied to the first and second groups. The mean Sen obtained with 5 ML algorithms applied to the four groups was 0.91. The highest Spe was 0.97 obtained with the LDA algorithm applied to the first and second groups. The mean Spe obtained with 5 ML algorithms applied to four groups was 0.91. The highest Mcc was 0.97 obtained with the LDA algorithm applied to the first group. The mean Mcc obtained with 5 ML algorithms applied to four groups was 0.91.
| Discussion | | |
Gender prediction determined with all or certain sections of the skeletal system is the first step in anthropological studies. Employment of the entire skeletal system often provides clear and accurate results; however, it is possible to achieve accurate results with a section of the skeletal system.[14] In the present study, a gender prediction test was conducted with an ML algorithm on CT images of 150 female and 150 male subjects without a pelvic skeleton pathology. The highest Acc was 0.96 with the LDA algorithm applied to the 1st group, and the highest Mcc was 0.97 with the LDA algorithm applied to the 1st group.
One of the key differences in the present study was the determination of the desired combinations of angles and lengths by placing landmarks on the pelvic skeleton. Thus, landmarks were placed on 35 anatomical locations on the pelvic skeleton and 34 lengths and 8 angles were obtained. In the literature review, it was observed that the highest number of parameters or angles and lengths were employed in the present study. d'Oliveira Coelho and Curate employed subpubic angle, pelvic height, and bispiniatic width, Pretorius et al. employed sciatic notch, Savall et al. employed 17 anatomical points on the os coxae, Blake et al. employed 10 anatomical locations on the pubis, Bonczarowska et al. employed various anatomical points on the ilium in gender prediction.[15],[16],[17],[18],[19] The present study employed the pelvis skeleton as a whole based on the lengths and angles obtained with the landmarks and contributed to the literature with a holistic approach.
In the present study, osteometry measurements were conducted on CT images instead of the measurements with conventional osteometry devices (calipers, odontometers, digital distance meters) employed for gender prediction. CT was preferred since it is a sensitive, modern method that is not affected by orientation and allows images to be transformed into the orthogonal form and reconstructed. CT imaging allowed the reconstruction of each bone section, providing an advantage in the measurement of missing and damaged sections.[2],[10]
The literature review demonstrated that orthogonalization was used in a few studies. In their study on the sternum, Oner et al. orthogonalized the sternum based on the T4 vertebra. In the present study, the images were superimposed and orthogonalized based on the promontory point on the pelvis skeleton. Thus, the images were not affected by orientation, and accurate and clear findings were obtained.[2]
Instead of conventional simple and advanced statistical methods, ML algorithms with higher reliability and accuracy were preferred. The literature review demonstrated that d'Oliveira Coelho and Curate reported 0.86 Acc for the pelvic skeleton with the RF algorithm.[17] In the present study, the RF algorithm revealed 0.92 Acc in the first group. In a study conducted on French, American, Thai, and Portuguese cranium bones, Santos et al. reported an Acc of 0.80 with LDA and LR algorithms.[20] In their study conducted on femoris, Curate et al. reported 0.85.5–92.5 Acc with LDA and 0.84–91 with LR.[11] In a study conducted on tarsal bones from the Coimbra skeleton collection, Navega et al. reported 0.86 Acc with LDA and LR.[21] In the present study, it was determined that Acc was 0.96 with the LDA algorithm in the first group, and 0.92 with the LR algorithm in the third group. In a study conducted on maxilla Akkoç et al. reported 0.90 Acc with RF algorithm.[22] In the present study, 0.92 Acc was determined with the RF algorithm in the first group. Based on previous study findings in the literature, this demonstrated that the pelvic skeleton was more dimorphic when compared to other bones.[3],[4]
The literature review revealed that the research samples employed in gender prediction with ML algorithms were smaller than the present study. The sample size was 40 in Akkoç et al., 194 in Bonczarowska et al., 113 in Savall et al., and 256 in d'Oliveira Coelho and Curate.[16],[17],[19],[22] The higher sample size in the present study and the employment of the ML algorithm significantly improved the performance criteria.
Mcc is a performance criterion that varies between −1 and +1. An Mcc of-1 indicates that all the classes are inversely predicted, an Mcc of 0 indicates that the results are random, and an Mcc of + 1 indicates that there is no error in the prediction. In the present study, an alternative Mcc was used to analyze Acc, Spe, and Sen, meaning that the reliability and accuracy were tested by various methods and reliable results were obtained in the study.[14]
The literature review demonstrated that gender prediction was never conducted with ETC and ADA algorithms before. In the present study, 0.93 Acc was obtained with the ETC algorithm in the second group, and 0.94 Acc was obtained with the ADA algorithm in the first and second groups. Thus, the present study employed an ML algorithm that was never used before, and it was demonstrated that it provided better results when compared to frequently used ML algorithms and other algorithms.
The measurements were performed on a specific population sample. For this reason, it may be more valuable to present data specific to each population. The fact that the measurements are repeated at different times and the images are brought to the orthogonal plane minimizes measurement errors. In our study, attention has been paid to these features to eliminate measurement errors. This can also be considered the strong side of our work.
| Conclusion | | |
According to the literature, high accuracy was obtained as a result of gender estimation with traditional osteometric measurements and basic statistics with modern imaging technologies and software-based ML algorithms. As seen in this study, the obtained accuracy rate increased up to 0.96. We believe that increasing the number of parameters will further increase the accuracy rate.
Financial support and sponsorship
Nil.
Conflicts of interest
There are no conflicts of interest.
| References | | |
| 1. | Colman KL, van der Merwe AE, Stull KE, Dobbe JG, Streekstra GJ, van Rijn RR, et al. The accuracy of 3D virtual bone models of the pelvis for morphological sex estimation. Int J Legal Med 2019;133:1853-60.  |
| 2. | Oner Z, Turan MK, Oner S, Secgin Y, Sahin B. Sex estimation using sternum part lenghts by means of artificial neural networks. Forensic Sci Int 2019;301:6-11. |
| 3. | Best KC, Garvin HM, Cabo LL. An investigation into the relationship between human cranial and pelvic sexual dimorphism. J Forensic Sci 2018;63:990-1000. |
| 4. | Spradley MK, Jantz RL. Sex estimation in forensic anthropology: Skull versus postcranial elements. J Forensic Sci 2011;56:289-96. |
| 5. | Pickering RB, Bachman DC. The Use of Forensic Anthropology. England: CRC Press; 1997. |
| 6. | Ali Z, Cox C, Stock MK, Zandee vanRilland EE, Rubio A, Fowler DR. Estimating sex using metric analysis of the scapula by postmortem computed tomography. J Forensic Sci 2018;63:1346-9. |
| 7. | Krems RV. Bayesian machine learning for quantum molecular dynamics. Phys Chem Chem Phys 2019;21:13392-410. |
| 8. | Park S, Park H, Im J, Yoo C, Rhee J, Lee B, et al. Delineation of high resolution climate regions over the Korean Peninsula using machine learning approaches. PLoS One 2019;14:e0223362. |
| 9. | Stoltzfus JC. Logistic regression: A brief primer. Acad Emerg Med 2011;18:1099-104. |
| 10. | Djorojevic M, Roldán C, García-Parra P, Alemán I, Botella M. Morphometric sex estimation from 3D computed tomography os coxae model and its validation in skeletal remains. Int J Legal Med 2014;128:879-88. |
| 11. | Curate F, Umbelino C, Perinha A, Nogueira C, Silva AM, Cunha E. Sex determination from the femur in Portuguese populations with classical and machine-learning classifiers. J Forensic Leg Med 2017;52:75-81. |
| 12. | Désir C, Petitjean C, Heutte L, Salaün M, Thiberville L. Classification of endomicroscopic images of the lung based on random subwindows and extra-trees. IEEE Trans Biomed Eng 2012;59:2677-83. |
| 13. | Gu W, Xie X, He Y, Zhang Z. Drug-target protein interaction prediction based on AdaBoost algorithm. Sheng Wu Yi Xue Gong Cheng Xue Za Zhi 2018;35:935-42. |
| 14. | Turan MK, Oner Z, Secgin Y, Oner S. A trial on artificial neural networks in predicting sex through bone length measurements on the first and fifth phalanges and metatarsals. Comput Biol Med 2019;115:103490. |
| 15. | Blake KA, Hartnett-McCann K. Metric assessment of the pubic bone using known and novel data points for sex estimation. J Forensic Sci 2018;63:1472-8. |
| 16. | Bonczarowska JH, Bonicelli A, Papadomanolakis A, Kranioti EF. The posterior portion of the ilium as a sex indicator: A validation study. Forensic Sci Int 2019;294:216.e1-6. |
| 17. | d'Oliveira Coelho J, Curate F. CADOES: An interactive machine-learning approach for sex estimation with the pelvis. Forensic Sci Int 2019;302:109873. |
| 18. | Pretorius E, Steyn M, Scholtz Y. Investigation into the usability of geometric morphometric analysis in assessment of sexual dimorphism. Am J Phys Anthropol 2006;129:64-70. |
| 19. | Savall F, Faruch-Bilfeld M, Dedouit F, Sans N, Rousseau H, Rougé D, et al. Metric sex determination of the human coxal bone on a virtual sample using decision trees. J Forensic Sci 2015;60:1395-400. |
| 20. | Santos F, Guyomarc'h P, Bruzek J. Statistical sex determination from craniometrics: Comparison of linear discriminant analysis, logistic regression, and support vector machines. Forensic Sci Int 2014;245:204.e1-8. |
| 21. | Navega D, Vicente R, Vieira DN, Ross AH, Cunha E. Sex estimation from the tarsal bones in a Portuguese sample: A machine learning approach. Int J Legal Med 2015;129:651-9. |
| 22. | Akkoç B, Arslan A, Kök H. Automatic gender determination from 3D digital maxillary tooth plaster models based on the random forest algorithm and discrete cosine transform. Comput Methods Programs Biomed 2017;143:59-65. |
[Figure 1], [Figure 2]
[Table 1], [Table 2], [Table 3], [Table 4], [Table 5], [Table 6]
|
Search Pubmed for