Saranya Honghimaphan, B.Sc.1, Patara Rattanachet, Ph.D.1, Natipong Chatthai, M.Sc.1, Parawee Jitrabeab, M.Sc.1, Sittiporn Ruengdit, Ph.D.2, Robert W.Mann, Ph.D.3, Napakorn Sangchay, M.D., Ph.D.1,*
1Siriraj Anatomical and Anthropological Bone Research Centre (Virapan Davivongs), Department of Anatomy, Faculty of Medicine Siriraj Hospital,
Mahidol University, Bangkok, Thailand, 2Department of Forensic Medicine, Faculty of Medicine, Chiang Mai University, Chiang Mai, Thailand, 3John
A. Burns School of Medicine, University of Hawai‘i, Honolulu, Hawai‘i, USA.
*Corresponding author: Napakorn Sangchay E-mail: napakorn.sac@mahidol.ac.th
Received 13 August 2025 Revised 10 October 2025 Accepted 18 October 2025 ORCID ID:http://orcid.org/0000-0001-9250-3702 https://doi.org/10.33192/smj.v77i12.277048
All material is licensed under terms of the Creative Commons Attribution 4.0 International (CC-BY-NC-ND 4.0) license unless otherwise stated.
ABSTRACT
Objective: This study investigates the sexual dimorphism in the patellae to develop sex estimation equations and determine the accuracy using machine learning (ML) and deep learning (DL) classification models.
Materials and Methods: The sample of 250 pairs of patellae from the Siriraj Anatomical and Anthropological Bone Research Center (Si-AABRC) were measured for eight parameters. The data were statistically analysed using logistic regression model alongside, ML and DL were used to predict the best classifiers in sex classification. Rather than traditional radiographic images, this paper tries a novel integration of photographic images.
Results: The average values for each parameter were significantly larger in males than females (p < 0.05), suggesting
the presence of sexual dimorphism within the patellae between each sex. The most dimorphic parameter was Transverse Diameter of Articular Facet (TDAF). The parameters in females showed no significant difference between left and right except for Breadth of the Medial Articular Facet (BMAF). However, in males a significant difference was observed for Maximum Height (MAXH), Transverse Diameter of Articular Facet (TDAF) and Breadth of the Lateral Articular Facet (BLAF). The logistic regression equation generated included the following parameters: MAXH (R), BLAF (L), and TDAF (L). The overall accuracy obtained for different sex estimation models ranged from 80%, 80% to 86% and 49.7% to 79.2% using logistic regression, ML and DL, respectively.
Conclusion: The patellae can be utilized by forensic anthropologists in determining the sex of an unknown individual.
Keywords: Patellae; sexual dimorphism; logistic regression; machine learning; deep learning; Thai population (Siriraj Med J 2025; 77: 829-846)
INTRODUCTION
Determining the identity of an unknown deceased is critical during a medicolegal investigation.1 A medical examiner and a forensic anthropologist work within the legal system to provide evidence and facilitate the identification of the deceased. This process ensures that justice is served to the individual and their family by resolving ambiguous questions, providing closure, and supplying accurate information for litigation purposes.2 In many instances, especially cold cases, the remains discovered are skeletonized. Hence, forensic anthropologists analyze the skeletal remains to establish a biological profile. The profile comprises sex, age, stature, and ancestry.3 Sex estimation is the initial step in the biological identification of the unknown as sex influences the other three parameters.4 Moreover, determining the sex minimizes the pool of candidates significantly by
50%.5
The human skeletal system is composed of several types of bones that can be categorized into long bones, flat bones, short bones, irregular bones and sesamoid bones. However, of all the bones in the human body, the bone that serve as the primary choice for anthropological sex estimation are a combination of irregular and flat bones; the pelvic (innominate) bone and the skull. The pelvic bone is primarily chosen for sex estimation as distinct morphological features can be witnessed between the sexes.6,7 Furthermore, the pelvis has been known to exhibit an accuracy of up to 95% in determining the
sex.8 Nonetheless, in the absence of the pelvic bone, the skull serves as the secondary choice in sex identification. This is because the accuracy has been observed to be approximately 92%.9 However, when unknown sets of remain are discovered, it is not always that a full set of skeletons is available. The incomplete nature of remains can be attributed to reasons such as animal scavenging, trauma, and the efficiency in retrieving and recovering the remains. Thus, in the context of forensic anthropology, different skeletal parts and bones are required to develop different methodologies in assessing the sex. One such bone is the patella.
The patella is the biggest sesamoid bone that is triangular in shape and possesses three borders – superior, medial and lateral, two surfaces – anterior and posterior, a base, and an apex.10 The patella is located within the patellofemoral groove of a femur and is embedded and attached to the quadricep femoris tendon and the patellar ligament.11 The posterior surface of the patella articulates with the femur forming an anterior articular surface of the knee joint that functions in protecting against trauma.2 Furthermore, the patella is also involved in the extensor mechanism of the quadricep muscle by acting as a fulcrum to increase the efficiency in the extension and flexion of the knee joint.12 Therefore, due to its morphology, location and function, the patella has been regarded as highly unsusceptible to post-mortem or taphonomy changes.13
In the past, patellae have been used in forensic cases to help identify unknown individuals. For example, in a 2008 case from Goias, Brazil,14 an unidentifiable decomposed body was discovered in a forest. Due to the severe state of decomposition, a fingerprint comparison was impossible. Alternatively, a positive identification was possible by comparing the patellae retrieved post-mortem to a radiographic image taken antemortem during knee surgery. The patella proved to be useful as the prior left knee surgery involving metallic surgical wires provided the evidence in the form of two channels in the patella that had exact dimensions to the surgical wires.14 Hence, this case showed that when a body is recovered, the conditions are not always favourable for identification. Several different parts of the skeleton should be used to determine the sex. In this instance, the patella has a location and position that allows it to resist post-mortem changes and is a useful tool in identifying the unknown. Thus, illustrates the importance of utilizing images to aid in distinguishing sex and identifying the individual Previously, numerous studies have assessed the sexual dimorphism of the patella. However, parameters regarded as the best predictors in sex estimation for both cranial and post cranial elements vary across populations, highlighting the importance of sexual dimorphism for each bone in specific population groups. While some studies have focused on morphometric measurments, others have incorporated machine learning for sex estimation.7,12 Nonetheless, many of these studies are limited by unequal15,16 or small sample sizes17,18, focus only on one side of the patella17 or restrict their method by primarily relying on discriminant function analysis for sex estimation.19 More recently, Cavlak et al20, have applied the framework of convolutional neural network to the MRI slices of patella, demonstrating the use of deep learning in sex estimation. As a result, building on these knowledge and foundation, this study aims to investigate sexual dimorphism using two different approaches; morphometric measurement inspected through logistic regression and machine learning, and photographic images evaluated through deep learning. In addition to addressing the comparative analysis, this studies aims to address limitation by using equal sample size and incorporating additional parameters. The contribution of this research is also in extending the image based analysis using photographs for a specific population. Therefore, it allows for a robust research by comparing traditional and advanced classification tools.
This is vital for forensic anthropologists to determine
the best method for estimating sex from the patella.
MATERIALS AND METHODS
This study analysed a total of 250 pairs of patellae in the Siriraj Anatomical and Anthropological Research Center’s (Si-AABRC) bone collection within the Department of Anatomy, Faculty of Medicine at Siriraj Hospital, Mahidol University. The sample consisted of an equal number of pairs of each sex, with an average age of 51 years. As both the left and right sides of the patella were used for the analysis, any remains without both patellae were excluded. Furthermore, patellae presented with fracture, erosion, trauma and extreme osteophytic formation were also excluded.
In this study, eight parameters in the patella were measured using a digital vernier calliper. The parameters were developed and adapted from the study conducted by Dayal & Bidmos19,21 assessing the sexual dimorphism of the patella in the South African population as well as the study assessing the sexual dimorphism of the patella in the Southern Part of Andhra Pradesh.17 The parameters shown in Fig 1. include maximum height (MAXH), maximum breadth (MAXB), maximum thickness (MAXT), height of the median ridge of the articular facet (HMRAF), breadth of lateral articular facet (BLAF), breadth of medial articular facet (BMAF), transverse diameter of articular facet (TDAF), and vertical diameter of the articular facet (VDAF). Detailed descriptions of the parameters are presented in Table 1. A patellar index (PI) was calculated to obtain a ratio between the MAXH and MAXL using the formula described below.
MAXH ×100
MAXB
The statistical analysis for the osteometric assessment of sexual dimorphism was conducted using Jamovi Version
2.5 - Sydney, Australia.22 The level of significance for each test was α = 0.05. Descriptive statistics comprising mean and standard deviation for all the parameters were calculated. The sample data were subjected to a Shapiro-Wilk normality test. If the data did not violate the assumption of normality, the difference between the left and right side of the patellae within each sex was assessed using independent t-test to determine the sexual dimorphism. However, if the data violated the aforementioned assumption, Welch’s t-test and Mann–Whitney U test were used, respectively. The percentage of sexual dimorphism was calculated using the formula
Fig 1. a) Patella measurements for the eight parameters. MAXH - Maximum Height, MAXB - Maximum Breadth, b) MAXT - Maximum Thickness, c) HMRAF - Height of median ridge of the articular facet, BLAF - Breadth of lateral articular facet, BMAF - Breadth of medial articular facet, d) VDAF - Vertical diameter of the articular facet, TDAF - Transverse diameter of the articular facet.
obtained from a study on sexual dimorphism of the root lengths in Tamil population.23
x
xmale - xfemale × 100
male
To assess the difference between the left and the right side of the patellae within each sex, the Wilcoxon Signed Rank (for data violating the normality) and paired t-test (for data following normal distribution) were used. Moreover, the mean difference was also calculated. With the presence of multiple parameters being analyzed, a Bonferroni and Benjamini–Hochberg test was also conducted and reported.
In determining the reliability of the parameters, 20% of the samples (25 pairs of males and 25 pairs
of females) were randomly selected for interobserver and intraobserver evaluations. The intraobserver and interobserver measurements were collected two weeks apart. The Intraclass Correlation Coefficient (ICC) was calculated to discover the degree of correlation and reliability between two observers.24
Following the guidelines by Koo and Yi (2016)25 and Shrout and Fleiss (1979)26, intraobserver reliability was assessed using ICC(3,1), a two-way mixed-effects model and interobserver reliability was analyzed using (ICC(2,k), a two-way random-effects model. Both the upper and lower confidence intervals was reported. The ICC score were interpreted as follows: ICC < 0.50 is considered poor, 0.50 – 0.75 is moderate, 0.75 – 0.90 is good and >0.90 is considered excellent.
In the next step, a logistic regression equation was
TABLE 1. Description of the parameters measured on the patellae.
Variable | Acronym | Description |
Maximum Height | MAXH | The maximum linear distance between the superior point of the apex to the inferior point at the base of the patellae |
Maximum Breadth | MAXB | The maximum linear distance between the medial border to the lateral border of the patellae |
Maximum Thickness | MAXT | The maximum linear distance between the anterior surface and the posterior surface of the patellae. |
Height of the Median Ridge of the Articular Facet | HMRAF | Distance between the superior point of the median ridge to the inferior point of the median ridge on the posterior surface of the patella |
Breadth of Medial Articular Facet | BMAF | Distance between the median ridge and the medial border of the articular facet on the posterior aspect of the patella. |
Breadth of Lateral Articular Facet | BLAF | Distance between the median ridge and the lateral border of the articular facet on the posterior aspect of the patella. |
Transverse Diameter of Articular Facet | TDAF | Maximum linear distance between the medial border of the articular facet and lateral border of the articular facet on the posterior aspect of the patella |
Vertical Diameter of Articular Facet | VDAF | Maximum linear distance between the most superior point of the articular facet and the most inferior point of the articular facet on the posterior aspect of the patella |
formulated using R Studio (Version 2024.4.2.764.1).27 The model was developed using the generalized linear model function – glm(). The output of this function represents a constant (a), a coefficient (b) and the measurement value of the predicting parameter (x) in the form of a + b × x.28 This model is then evaluated for the log of odd ratio (logit) and a probability of classification equation (P).
1
1+e-(a+b1.x1+b2.x2. )
The equation for the probability of classification predicts the outcome to be either male or female, where P > 0.5 classifies the measurement to be male and P ≤ 0.5 classifies the measurement to be female.29 A logit equation for each parameter and each side was developed. However, as there are multiple predictors, a stepwise model selection function was performed to obtain the parameter that provides the best efficacy for sex estimation based on an AIC, which is a computational value that depicts the efficacy of the model.28 The model that presented with the lowest AIC and highest Area Under the Curve (AUC) was selected. A ROC Curve as well as a Calibration plot was generated. For the model that provides the best efficacy,
a Hosmer–Lemeshow χ² alongside p-value and a Brier score was calculated. The model was then tested using a blind test where 100 additional pairs of the patellae were measured and the values were put in the model to determine the accuracy.
Machine Learning analyses were performed using MATLAB version 9.14 - 2023a30 and Waikato Environment for Knowledge Analysis software (WEKA 3.8.6).31 To prevent the subject-level leakage, each row in the dataset represented one subject containing the measurement of both left and right patella.
For analysis in WEKA, 250 subject pairs were randomly assigned to 10 groups, as the models were evaluated using 10-fold cross validation. During the training phase, the models were cross-validated using a separate test dataset selected through the ‘Supplied Test Set’ option. In each fold models were trained on 90% of the dataset and validated on the 10% of the dataset. The mean and standard deviation of the performance metrics were calculated across 10 folds.
For analysis in MATLAB, the response variable was set at “Sex”. The model training was then conducted
using the Classification Learner application with 10 fold cross validation. The mean and standard deviation of the performance metrics were reported for models exhibiting more than 80% of accuracy.
For the deep learning approach, 1000 images [left, right, anterior and posterior] were photographed using an iPhone 15 Pro at 3x magnification placed on an instrument (Fig 2) at a height of 24 cm. Each patella was secured with a clay to ensure a levelled orientation allowing the anterior surface and the posterior surface of the patella to be photographed (Fig 3). The deep learning assessment was done to only classify using anterior surface and only classify using posterior surface
Images in the dataset of the left and right patella
from the same subject were assigned exclusively to either training set (“Training → Male/Female”) or the validation test (“Test → Male/Female”) for both anterior images and posterior images, ensuring a subject-level split. Pretrained models like Visual Geometry Group (VGG19), MobileNet V2 and Deep Residual Learning (ResNet18) was utilized. The input images were resized to 224 pixels x 224 pixels with single precision and processed to 3 channel RGB. An augmentation was applied including random rotation (±10°), translation (±10 pixels), and scaling (0.9–1.1) in both axes.
Following this a 5-fold cross validation was utilized on the training dataset where each were set for 20 epochs, initial learning rate = 1 × 10-⁴ and batch size of 32. The sample was evaluated on a test set. Performance metrics was evaluated; accuracy, sensitivity and specificity.
a)
b)
Fig 2. a) A detailed schematic created using the Canva software showcasing the instrument used in taking images of the patellae. b) Images showcasing the set-up used for photographing the patella.
a) b)
Fig 3. a) Representative image of the anterior view of the patella. b) Representative image of the posterior view of the patella
RESULTS
Sexual dimorphism assessment and descriptive statistics Descriptive statistics consisting of mean and standard deviation (SD) were determined for all eight parameters for both sexes and sides. Comparing each parameter between males and females using parametric and non-parametric test, a p < 0.001 which is less than α = 0.05 was obtained. Seen in Table 2, the percentage of sexual dimorphism for each parameter is positive. The highest percentage and lowest percentage of sexual dimorphism were observed in TDAF (Left) at 12.15% and HMRAF (Right) at 6.72% respectively. Furthermore, the average PI on the left side of male was 95.97 and female was
95.82. The average patellar PI on the right side of both male and female were slightly smaller at 95.23 and 95.42 respectively. This is concurrent with the Bonferroni and Benjamini–Hochberg test where all parameters were significant. (q = 0.002) (Table 3).
Seen in Table 3, the results assessing the difference in the patellar side for each parameter for both males and females are summarized. All the parameters in females only the BMAF had a significant difference (p < 0.001), whereas in males MAXH (p < 0.001), TDAF (p = 0.041), and BLAF (p = 0.012) had a significant difference. However, from performing the Bonferroni and Benjamini–Hochberg only MAXH (q = 0.002) and BLAF (q = 0.020) showed significant difference within males and for females only BMAF (q = 0.002) (Table 3).
Reliability analysis by interobserver and intraobserver The ICC values for intraobserver and interobserver measurements on 50 randomly selected samples are summarized in Table 4. The intraobserver values ranged from 0.842 to 0.997 and the interobserver values ranged
from 0.502 to 0.995.
Regarding the both the reliability analysis, it was found that the highest ICC value belonged to the MAXB in both males and females. For the intraobserver value, the highest ICC was 0.995 for male on the left side and 0.997 for female on the right side. For the interobserver value, the highest ICC was 0.995 for male on the right side and 0.994 for both sides in female.
However, the parameter that proved to show the lowest reliability in interobserver measurement was the VDAF for both male and female as well as for both sides with values ranging from 0.502 to 0.706. The parameters that showcased the highest reliability for both male and female include MAXH, MAXB, and MAXT.
The logistic regression analysis as seen in Table 5 indicates TDAF–transverse diameter of the articular facet on the left side as the most effective parameter in sex determination. This is because the model for TDAF produced the lowest AIC score of 212.20 and the highest AUC of 0.8941 (Table 5). Additionally, the ROC curve presented in Fig 5 visually demonstrates differences between the parameters, highlighting TDAF on the left side taking precedence in sex estimation.
TABLE 2. Descriptive statistics and percentage of sexual dimorphism of each parameter for both the sexes and both the sides for the patellae.
Male | Female | Benjamini | Bonferroni | ||||||
Parameters | (n=125) | (n=125) | % Sexual P | Hochberg | |||||
and Side | Mean | SD | Mean | SD | Dimorphism Value | P (All) | P (Within) | P (All) | P (Within) |
MAXH (L)a | 42.2 | 3.41 | 37.3 | 2.67 | 11.61% <0.001 | 0.00178 | 0.001 | 0.032 | 0.008 |
MAXH (R)a | 41.7 | 3.27 | 37.2 | 2.61 | 10.79% <0.001 | 0.00178 | 0.001 | 0.032 | 0.008 |
MAXB (L) | 44.0 | 3.18 | 39.0 | 3.01 | 11.36% <0.001 | 0.00178 | 0.001 | 0.032 | 0.008 |
MAXB (R) | 43.9 | 3.20 | 39.1 | 3.09 | 10.93% <0.001 | 0.00178 | 0.001 | 0.032 | 0.008 |
MAXT (L)b | 20.8 | 1.80 | 18.7 | 1.49 | 10.10% <0.001 | 0.00178 | 0.001 | 0.032 | 0.008 |
MAXT (R)b | 20.7 | 1.80 | 18.6 | 1.49 | 10.14% <0.001 | 0.00178 | 0.001 | 0.032 | 0.008 |
HMRAF (L) | 26.7 | 2.91 | 24.8 | 2.37 | 7.12% <0.001 | 0.00178 | 0.001 | 0.032 | 0.008 |
HMRAF (R) | 26.8 | 2.79 | 25.0 | 2.41 | 6.72% <0.001 | 0.00178 | 0.001 | 0.032 | 0.008 |
BMAF (L)b | 20.4 | 2.15 | 18.1 | 1.85 | 11.27% <0.001 | 0.00178 | 0.001 | 0.032 | 0.008 |
BMAF (R) | 20.6 | 1.89 | 18.6 | 1.88 | 9.71% <0.001 | 0.00178 | 0.001 | 0.032 | 0.008 |
BLAF (L) | 27.0 | 2.04 | 23.9 | 1.90 | 11.48% <0.001 | 0.00178 | 0.001 | 0.032 | 0.008 |
BLAF (R)b | 26.7 | 2.17 | 23.8 | 1.96 | 10.86% <0.001 | 0.00178 | 0.001 | 0.032 | 0.008 |
TDAF (L) | 42.8 | 3.09 | 37.6 | 2.97 | 12.15% <0.001 | 0.00178 | 0.001 | 0.032 | 0.008 |
TDAF (R) | 42.5 | 3.14 | 37.6 | 2.97 | 11.53% <0.001 | 0.00178 | 0.001 | 0.032 | 0.008 |
VDAF (L)b | 31.6 | 2.74 | 29.1 | 2.46 | 7.91% <0.001 | 0.00178 | 0.001 | 0.032 | 0.008 |
VDAF (R) | 31.5 | 2.35 | 29.0 | 2.27 | 7.94% <0.001 | 0.00178 | 0.001 | 0.032 | 0.008 |
n = Sample size (pair of patellae)
a = Levene’s test (Homogeneity of variance test): violation of the assumption of equal variances. Welch’s t-test.
b = Shapiro Wilk (Normality Test): violation of the assumption of normality. Mann-Whitney U t-test.
Abbreviations: MAXH = Maximum Height, MAXB = Maximum Breadth, MAXT = Maximum Thickness, HMRAF = Height of the Median Ridge of the Articular Facet, BMAF = Breadth of Medial Articular Facet, BLAF = Breadth of Lateral Articular Facet, TDAF = Transverse Diameter of Articular Facet, VDAF = Vertical Diameter of Articular Facet.
Parameters | Mean Difference | P Value | Male (n=125) Benjamini Hochberg | Bonferroni | Mean Difference | Female (n=125) P Benjamini Value Hochberg | Bonferroni | |||||
P (All) | P (Within) | P (All) | P (Within) | P (All) | P (Within) | P (All) | P (Within) | |||||
MAXH (L&R) | 0.5100a | <0.001 | 0.002 | 0.008 | 0.032 | 0.008 | 0.1128 | 0.207 | 0.276 | 0.552 | 1.000 | 1.000 |
MAXB (L&R) | 0.1329 | 0.24 | 0.307 | 0.320 | 1.000 | 1.000 | -0.07 | 0.862 | 0.890 | 0.862 | 1.000 | 1.000 |
MAXT (L&R) | 0.0400a | 0.400a | 0.457 | 0.457 | 1.000 | 1.000 | 0.01 | 0.591 | 0.652 | 0.788 | 1.000 | 1.000 |
HMRAF(L&R) | -0.2000a | 0.064a | 0.098 | 0.128 | 1.000 | 0.512 | -0.1400a | 0.113a | 0.157 | 0.452 | 1.000 | 0.904 |
BMAF (L&R) | -0.1700a | 0.098a | 0.143 | 0.157 | 1.000 | 0.784 | -0.4501a | <0.001a | 0.002 | 0.008 | 0.032 | 0.008 |
BLAF (L&R) | 0.2650a | 0.012a | 0.020 | 0.048 | 0.384 | 0.096 | 0.0549a | 0.388a | 0.457 | 0.621 | 1.000 | 1.000 |
TDAF (L&R) | 0.2600a | 0.041a | 0.066 | 0.109 | 1.000 | 0.328 | -0.0450a | 0.824a | 0.879 | 0.862 | 1.000 | 1.000 |
VDAF (L&R) | 0.0150a | 0.890a | 0.890 | 0.890 | 1.000 | 1.000 | 0.0500a | 0.333a | 0.410 | 0.621 | 1.000 | 1.000 |
TABLE 3. Comparison between the left and right sides for each parameter, separately for males and females.
a
L = Left and R = Right.
n = Sample size (pair of patellae)
a = Shapiro Wilk (Normality Test): violation of the assumption of normality. Wilcoxon signed-rank test
Abbreviations: MAXH = Maximum Height, MAXB = Maximum Breadth, MAXT = Maximum Thickness, HMRAF = Height of the Median Ridge of the Articular Facet, BMAF = Breadth of Medial Articular Facet, BLAF = Breadth of Lateral Articular Facet, TDAF = Transverse Diameter of Articular Facet, VDAF = Vertical Diameter of Articular Facet.
TABLE 4. Intraclass Correlation Coefficient (ICC) and descriptive analysis for the Intraobserver and Interobserver measurement.
Intraobserver (n=25) ICC(3,1) [95% CI]
MAXH L 42.5 4.31 0.981 0.962 0.990 37.1 2.17 0.994 0.989 0.997
MAXB
L
R
3.92 0.995 0.990 0.998 39.0 2.89 0.996 0.993
3.97 0.993 0.987 0.997 39.3 2.73 0.997 0.994
0.998
0.998
R 42.1 4.41 0.984 0.969 0.992 37.3 2.13 0.993 0.986 0.996
MAXT L 20.3 1.93 0.986 0.973 0.993 18.5 1.30 0.982 0.965 0.991
HMRAF
L
R
26.6 3.22 0.916 0.840 0.957 24.8 2.24 0.873 0.766
27.2 2.94 0.951 0.906 0.975 24.7 2.42 0.903 0.819
0.933
0.949
R 20.1 1.77 0.992 0.984 0.996 18.6 1.28 0.957 0.917 0.978
BMAF L 20.8 1.77 0.861 0.741 0.927 18.3 1.78 0.894 0.802 0.944
BLAF
L
R
26.5 2.26 0.927 0.859 0.962 23.9 1.77 0.863 0.749
26.4 2.78 0.960 0.922 0.979 23.8 1.79 0.842 0.713
0.928
0.916
R 20.9 1.83 0.906 0.821 0.951 19.1 1.58 0.852 0.729 0.921
TDAF L 42.7 3.74 0.982 0.965 0.991 37.6 2.62 0.973 0.948 0.986
R 42.5 3.69 0.980 0.961 0.990 37.7 2.36 0.938 0.881 0.968
VDAF | L R | 31.0 31.7 | 2.65 2.58 | 0.967 0.969 | 0.935 0.940 | 0.983 0.984 | 29.2 29.1 | 1.89 1.98 | 0.965 0.967 | 0.933 0.937 | 0.982 0.983 |
Interobserver (n=25) ICC(2,k) [95% CI] | |||||||||||
MAXH | L R | 42.0 41.3 | 4.36 4.63 | 0.984 0.984 | 0.967 0.955 | 0.992 0.993 | 36.4 36.4 | 2.16 2.22 | 0.948 0.926 | 0.605 0.391 | 0.983 0.976 |
MAXB L 43.5 3.76 0.992 0.977 0.996 38.8 2.85 0.994 0.981 0.997
MAXT
L
R
19.9 1.91 0.973 0.830 0.991 18.3 1.31 0.980 0.895
19.8 1.81 0.970 0.890 0.988 18.4 1.38 0.956 0.786
0.992
0.983
R 43.7 3.90 0.995 0.988 0.998 39.1 2.60 0.994 0.973 0.998
HMRAF L 25.3 3.46 0.867 0.462 0.949 23.6 2.05 0.781 0.128 0.917
BMAF
L
R
19.6 1.91 0.874 0.499 0.951 17.4 1.75 0.898 0.406
19.5 1.71 0.791 0.158 0.921 18.2 1.57 0.820 0.383
0.964
0.927
R 25.4 3.26 0.876 0.036 0.963 23.6 2.29 0.786 0.220 0.915
BLAF L 25.9 2.54 0.911 0.588 0.967 23.6 1.92 0.889 0.472 0.959
TDAF
L
R
41.4 3.70 0.951 0.571 0.984 36.5 2.57 0.910 0.110
41.0 3.49 0.923 0.262 0.976 37.1 2.69 0.931 0.411
0.973
0.978
R 25.8 2.73 0.949 0.563 0.984 23.6 1.91 0.911 0.819 0.955
VDAF L 28.5 2.35 0.706 -0.112 0.898 26.6 1.90 0.502 -0.142 0.791
R 28.8 2.23 0.665 -0.125 0.879 26.6 2.0 0.586 -0.140 0.839
a = Intraclass Correlation Coefficient for the comparison between the intraobserver measurement and original as well as, interobserver measurement and original
Abbreviations: MAXH = Maximum Height, MAXB = Maximum Breadth, MAXT = Maximum Thickness, HMRAF = Height of the Median
Ridge of the Articular Facet, BMAF = Breadth of Medial Articular Facet, BLAF = Breadth of Lateral Articular Facet, TDAF = Transverse Diameter of Articular Facet, VDAF = Vertical Diameter of Articular Facet.
TABLE 5. Logistics Regression Model for single and combination of parameters for sex estimation.
Single Logistic Regression
MAXH L -19.993 + 0.505 ×MAXHL 230.67 1.66 (1.48 – 1.89) 0.8652
MAXB
L
R
-21.415 + 0.516×MAXBL
-19.701 + 0.475×MAXBR
227.33
237.55
1.68 (1.49 – 1.92)
1.61 (1.44 – 1.83)
0.8786
0.8634
R -20.120 + 0.512×MAXHR 236.92 1.67 (1.48 – 1.91) 0.8607
MAXT L -15.894 + 0.809×MAXTL 263.66 2.25 (1.83 – 2.83) 0.8210
HMRAF
L
R
-7.190 + 0.280×HMRAFL
-7.030 + 0.272×HMRAFR
319.27
321.28
1.32 (1.19 – 1.48)
1.31 (1.18 – 1.47)
0.6977
0.6921
R -15.497 + 0.791×MAXTR 267.27 2.21 (1.80 – 2.77) 0.8190
BMAF L -11.445 + 0.594×BMAFL 276.02 1.81 (1.55 – 2.16) 0.8062
BLAF
L
R
-20.209 + 0.795×BLAFL
-17.549 + 0.697×BLAFR
232.64
250.69
2.22 (1.84 – 2.74)
2.01 (1.70 – 2.43)
0.8709
0.8458
R -11.089 + 0.566×BMAFR 286.98 1.76 (1.51 – 2.09) 0.7835
TDAF L -22.917 + 0.571×TDAFL 212.20 1.77 (1.56 – 2.05) 0.8941
R -20.223 + 0.505×TDAFR 229.50 1.66 (1.48 – 1.89) 0.8767
VDAF | L R | -12.600 + 0.416×VDAFL -13.677 + 0.453×VDAFR | 296.40 1.52 (1.34 – 1.74) 288.44 1.57 (1.39 – 1.81) | 0.7694 0.7784 |
Best Logistic Regression Model | ||||
Predictor | Coefficient | Std. Error P-value | Odds Ratio (95% CI) AIC | AUC |
Intercept | -26.87 | 3.30 | <0.001 | 2.14 x 10-12 | ||
(1.92 x 1015 – 8.50 x 1010) | 0.907 | |||||
MAXH.R | 0.219 | 0082 | 0.007 | 1.245 (1.065 – 1.469) | 205 | (95% CI: |
BLAF.L | 0.281 | 0.149 | 0.059 | 1.325 (0.993 – 1.783) | 0.8685-0.945) | |
TDAF.L | 0.277 | 0.110 | 0.012 | 1.319 (1.070 – 1.652) | ||
Threshold Metrics of the Best Logistic Regression Model | ||||||
Sensitivity | Specificity | Precision (PPV) | NPV | Accuracy (95% CI) | ||
0.864 | 0.808 | 0.818 | 0.856 | 0.836 | ||
(0.7842 - 0.8797) | ||||||
Blind Test (n=100) on the Best Logistic Regression Model | ||||||
0.76 0.84 0.79 0.81 0.80
Fig 4. A confusion matrix representing the performance evaluation of the best model obtained through logistic regression analysis.
Fig 5. Receiver operating characteristic curve (ROC curves) for the all the univariable measured on the patella.
Moreover through stepwise logistic regression, the most notable model for sex estimation incorporated the parameters MAXH(R), BLAF(L), TDAF(L), formulating the equation: -26.87+(TDAF.L×0.277)+(MAXH.
R×0.219)+(BLAF.L×0.281) (Table 5). However, within these three the parameter that contributes most to the prediction is MAXH(R) and TDAF(L) as their p values are statistically significant at 0.007 and 0.012 respectively. The analysis resulted in an AIC score of 205 and an AUC of 0.907 (95% CI: 0.8685 - 0.945) (Fig 6). The performance metrics revealed a sensitivity of 0.864, specificity of 0.808,
a precision of 0.818, NPV of 0.856 and accuracy of 0.836 (95% CI: 0.7842 - 0.8797) (Table 5). A Hosmer–Lemeshow test revealed X2 of 8.2406 with a p-value of 0.4103 – an appropriate goodness of fit. In addition, the Brier score revealed a low prediction error with a value of 0.12.
Different models were obtained from using WEKA software and MATLAB software that exhibited the accuracy of 80% or more (Table 6). The models in MATLAB that produced the highest were Coarse Tree and Coarse
a)
b)
c)
Fig 6. a) Receiver operating characteristic curve (ROC curves) for the equation that best determines sex in patella b) Receiver operating characteristic curve (ROC curves) with confidence band. c) Calibration plot for the equation that best determines sex in patella: TDAF.L + MAXH.R + BLAF.L.
TABLE 6. Comparing the top ten different machine learning models from WEKA and MATLAB to determine the model with the highest accuracy in sex estimation.
WEKA | ||||||
Model | Accuracy (%) | TP Rate Recall Sensitivity | FP Rate | Precision PPV | F-measure | Specificity |
Bayes Net | 83.60±0.06 | 0.85±0.06 | 0.17±0.11 | 0.84±0.10 | 0.84 ±0.06 | 0.83±0.11 |
Naives Bayes | 83.2±0.07 | 0.83±0.09 | 0.17±0.10 | 0.84±0.09 | 0.83±0.07 | 0.83±0.10 |
Logistic | 81.2±0.08 | 0.82±0.10 | 0.19±0.12 | 0.82±0.11 | 0.81±0.08 | 0.81±0.12 |
SGD | 82±0.07 | 0.82±0.09 | 0.18±0.12 | 0.83±0.11 | 0.82±0.07 | 0.82±0.12 |
Simple Logistic | 82.4±0.09 | 0.85±0.07 | 0.2±0.14 | 0.82±0.12 | 0.83±0.08 | 0.8±0.14 |
SMO | 84.4±0.05 | 0.84±0.07 | 0.15±0.10 | 0.86±0.09 | 0.84±0.04 | 0.85±0.10 |
LWL | 84.4±0.08 | 0.8±0.10 | 0.11±0.09 | 0.88±0.09 | 0.84±0.09 | 0.89±0.09 |
Decision Tree | 82.8±0.08 | 0.83±0.10 | 0.17±0.09 | 0.83±0.08 | 0.83±0.08 | 0.83±0.09 |
JRip | 86±0.09 | 0.85±0.10 | 0.13±0.12 | 0.88±0.11 | 0.86±0.09 | 0.87±0.12 |
OneR | 86±0.08 | 0.87±0.11 | 0.15±0.12 | 0.86±0.08 | 0.86±0.07 | 0.85±0.12 |
PART | 82.4±0.08 | 0.78±0.07 | 0.13±0.12 | 0.87±0.11 | 0.82±0.08 | 0.87±0.12 |
Decision Stump | 84.4±0.08 | 0.8±0.10 | 0.11±0.09 | 0.88±0.09 | 0.84±0.09 | 0.89±0.09 |
Hoeffding Tree | 83.2±0.07 | 0.82±0.09 | 0.16±0.10 | 0.84±0.09 | 0.83±0.07 | 0.84±0.10 |
LMT | 82.8±0.08 | 0.83±0.09 | 0.18±0.11 | 0.83±0.10 | 0.83±0.08 | 0.83±0.11 |
Random Forest | 82±0.07 | 0.8±0.08 | 0.16±0.11 | 0.84±0.10 | 0.82±0.07 | 0.84±0.11 |
REP Tree | 83.2±0.08 | 0.82±0.09 | 0.16±0.12 | 0.85±0.10 | 0.83±0.07 | 0.84±0.12 |
AdaBoostM1 | 84±0.08 | 0.81±0.09 | 0.13±0.08 | 0.86±0.08 | 0.83±0.08 | 0.87±0.08 |
Attribute Selected Classifier | 83.6±0.06 | 0.79±0.06 | 0.12±0.10 | 0.87±0.10 | 0.83±0.07 | 0.88±0.10 |
Bagging | 84.4±0.05 | 0.83±0.06 | 0.14±0.07 | 0.86±0.07 | 0.84±0.05 | 0.86±0.07 |
Classification via Regression | 84.4±0.08 | 0.8±0.09 | 0.11±0.09 | 0.88±0.09 | 0.84±0.08 | 0.89±0.09 |
Filtered Classifier | 82.4±0.06 | 0.8±0.06 | 0.15±0.08 | 0.84±0.08 | 0.82±0.06 | 0.85±0.08 |
Iterative Classifier Optimizer | 82.8±0.06 | 0.79±0.08 | 0.13±0.09 | 0.86±0.09 | 0.82±0.07 | 0.87±0.09 |
Logit Boost | 82.8±0.06 | 0.79±0.08 | 0.14±0.07 | 0.86±0.07 | 0.82±0.07 | 0.86±0.07 |
Random Subpace | 82.8±0.08 | 0.82±0.08 | 0.16±0.13 | 0.85±0.12 | 0.83±0.08 | 0.84±0.13 |
MATLAB | ||||||
Coarse Tree | 83.7±0.02 | 0.81±0.01 | 0.13±0.02 | 0.86±0.02 | 0.83±0.02 | 0.87±0.02 |
Linear Discriminant | 81.7±0.01 | 0.81±0.02 | 0.18±0.01 | 0.82 | 0.82±0.0 | 0.82±0.01 |
Binary GLM LR | 80.9 | 0.81±0.01 | 0.19±0.01 | 0.81±0.01 | 0.81±0. | 0.81±0.01 |
Efficient LR | 81.2±0.01 | 0.82±0.01 | 0.19 | 0.81 | 0.81±0.01 | 0.81 |
Gaussian Naive Bayes | 83.2±0.01 | 0.82±0.01 | 0.16 | 0.84±0.01 | 0.83±0.01 | 0.84±0.01 |
Kernel Naive Bayes | 83.3±0.01 | 0.83 | 0.16±0.01 | 0.84±0.01 | 0.83±0.01 | 0.84±0.01 |
Linear SVM | 82.7±0.01 | 0.83±0.01 | 0.18±0.01 | 0.82±0.01 | 0.83±0.01 | 0.82±0.01 |
Medium Gaussian SVM | 81.9±0.01 | 0.80 | 0.16±0.01 | 0.83±0.01 | 0.81±0.01 | 0.84±0.01 |
Coarse Gaussian SVM | 83.7 | 0.84±0.01 | 0.17±0.01 | 0.83±0.01 | 0.84±0.01 | 0.83±0.01 |
Medium KNN | 80.7±0.01 | 0.78±0.01 | 0.17±0.01 | 0.82±0.01 | 0.80±0.01 | 0.83±0.01 |
Coarse KNN | 82.4±0.01 | 0.80±0.01 | 0.15±0.01 | 0.84±0.01 | 0.82±0.01 | 0.85±0.01 |
Cosine KNN | 82.8 | 0.82 | 0.17±0.01 | 0.83 | 0.83 | 0.83±0.01 |
Cubic KNN | 80.9±0.01 | 0.79 | 0.17±0.01 | 0.82±0.01 | 0.80±0.01 | 0.83±0.01 |
Weighted KNN | 80.4±0.01 | 0.82±0.01 | 0.21±0.03 | 0.80±0.02 | 0.81±0.02 | 0.79±0.03 |
Bagged Tree | 82.5±0.01 | 0.79±0.02 | 0.14±0.01 | 0.85±0.01 | 0.82±0.02 | 0.86±0.01 |
Subspace Discriminant | 83.3±0.01 | 0.83±0.02 | 0.17±0.01 | 0.83±0.01 | 0.83±0.01 | 0.83±0.01 |
Subspace KNN | 80.7±0.01 | 0.79±0.01 | 0.18±0.02 | 0.82±0.02 | 0.80±0.01 | 0.82±0.02 |
TABLE 7. Deep Learning Models classifying male and female patellae.
Model | Cross Validation | Test | ||||
Accuracy (%) | Sensitivity | Specificity | Accuracy (%) | Sensitivity | Specificity | |
VGG19 | 50.6±0.13 | 0.30±0.27 | 0.2±0.27 | 49.7 | 0 | 1 |
ResNet18 | 79.2±0.04 | 0.64±0.12 | 0.94±0.03 | 79.0 | 0.65 | 0.94 |
MobileV2 | 65.2±0.05 | 0.55±0.24 | 0.76±0.32 | 53.8 | 0.335 | 0.74 |
Gaussian SVM at 83.7% while the lowest was Weighted KNN at 80.4%. On the other hand, in WEKA the model with the highest validation accuracy is Jrip and OneR with 86.0% and the lowest accuracy is Logistic (81.2%). Hence, both WEKA and MATLAB provide models with similar ranging validation accuracy.
In addition, from the three models that were trained and tested in classification of images generated the best results with the accuracy of 79% for the ResNet18. Whereas the lowest accuracy obtained was for VGG19 with an accuracy of 49.7%.
DISCUSSION
The aim of this investigation was to conduct a comparative analysis of several sex estimation modalites and evaluate the efficacy of deep learning model to machine learning and logistic regression for sex estimation using the dry patellae. By combining morphometric and image-based analysis, this paper extends traditional statistical models and discriminant functions. The results obtained help reinforce that patella showcases apparent sexual dimorphism and can provide accuracy in the absence of the primary skeleton choice (pelvis and skull).
Furthermore, the advantage of using patella can be attributed to both the external factors and internal factors like hormones, genes and biomechanics that influence the osseous robustivity within each sex.32 Therefore, other bones such as the patella can be of an alternative. The anatomical position, dense33 composition and the articular cartilage of the patellae is composed of copious amounts of collagen and glycosaminoglycan34, making it resilient to taphonomic changes. As a result it increases the probability of recovery during excavation.
the patella were larger in males than female, attributing to the hormonal factor like oestrogen which has a stimulatory effect to the increase in the periosteal growth in males but contrary in females, acts as inhibitory effect on the periosteal growth.35 These results are in accordance with the findings of several previous researches conducted on different populations such as South African Whites19, South African of European ancestry18, Central Thai population10, Northern Thai population15, Spanish population13, Southern Part of Andhra Pradesh, India17 and Southern Italian population.36
Comparing the average between sexes within regional boundaries, the patella in this study exhibited a larger value for the HMRAF and BMAF in comparison to the Northern Thai region15 and Central Thai region.10 The hypothesis for this could be variation within a biology of a subgroup of a population but also it could be the indefinite landmark on the posterior aspect of the patella. Furthermore, comparing the mean values of this study to the findings from Southern Chinese population37 and Korean population (total knee arthroplasty)38 can infer that Asian population may have smaller patellar dimensions compared to Westerners, although there has also been a study on the Asian population with healthy knees that revealed no difference between Asians and Westerners.39 The discrepancy in the measurement can be attributed to to the different genetic makeup, lifestyle and environment.15
In the forensic context, the differences in these
datahighlight the importance of using population specific reference. Furthermore, with the presence of apparent sexual dimorphism, the patella could be used for sex estimation in the field of forensic anthropology. In the area where remains are found, if the primary bones are not available, patella can be the supplementary element to help determine the sex.
Furthermore, most of the findings in our study were consistent to both sides of the patella, indicating that any
side can be used in sex estimation. However, minor side specific difference was observed in BMAF for females and MAXH and BLAF in males. While these finding contradicts the information from previous research on Portuguese population16 and Spanish population13 reporting no significant difference, one plausible reason could be the existence of bilateral variation. Nevertheless, the differences in skeletal measurements may be limited or inconsistent across different parameters, suggesting a minimal impact. In addition, the variation of these significant differences between left and right side lacks comparative studies on the bilateral factor of patella, suggesting a need for further research.
Reliability analysis by interobserver and intraobserver The ICC values for the interobserver and intraobserver measurement ranged from 0.502 to 0.995 and 0.842 to 0.997 respectively. The parameters that showcased the highest reliability for both male and female include MAXH, MAXB, and MAXT because these parameters have definite landmarks with methodology being repeated by many authors. In addition, research on the Thai population by Srinak and Sukvitchai10 reported similar results. This reinforces the repeatability of the methodology being used. However, the lowest interobserver measurement can be attributed to the idea that even though the observers were using the same description, the inconsistency and subjectivity of the landmarks can cause the discrepancies. The VDAF parameter had the lowest value and this could be due to the posterior side of the patella having a smooth surface. Therefore, the margins are not as defined compared to the other parameters. Moreover, this also further suggest that parameters such as VDAF should probably be omitted when estimation sex of an
unidentified individual.
In the field of forensic science, two of the most common methods that are used in formulating a sex estimation equation are logistic regression and discriminant function analysis (DFA).41 Thus, unlike most of the previous studies on sexual dimorphism on the patella that uses DFA, this study performed a logistic regression analysis to evaluate the parameters that estimate sex. The reason for this is that the data obtained in this study have parameters with measurements that are not normally distributed. Hence, logistic regression is more suitable because it can provide robust estimations even when assumptions are violated.41 Additionally, logistic regression offers greater flexibility compared to DFA because the generated equation can be directly utilized
to compute the probabilities of categorizing the sexes, thereby quantifying uncertainties. In contrast, DFA solely provides the equation and cut-off point, resulting in the absence of prior probabilities. This absence diminishes the confidence in classifying the sexes, potentially leading to misclassification.42
Many previous studies have reported that the measurement that exhibited the highest sexual dimorphism varied among populations as it included MAXH (South African White, Portuguese, Spanish)13,16,19 MAXB (Portuguese)16, MAXT (Italian and Spanish)13,36, BLAF (Central Thai and Northern Thai).10,15 However, in this study, TDAF was most sexually dimorphic with 12.15% on the left side and 11.53% on the right side. This is further iterated by the logistic regression analysis, producing the highest AUC but the lowest AIC value. TDAF is not a common parameter measured in previous studies but rather divided into the breadth of medial facet (BMAF) and breadth of lateral facet (BLAF). Furthermore, a stepwise regression analysis revealed the parameters identified as the most notable for sex estimation were MAXH.R, BLAF.L, and TDAF.L. The performance metrics for this LR model revealed a strong predictability model. With high sensitivity and specificity value as well as precision and negative predictive value, the model is very effective in accurately distinguishing between the sexes. A blind test of 100 samples revealed an accuracy 80.0%. Thus, these results are comparable with other populations such as the 73% - 89%.10,13,15,19,21 The value obtained in this study is in the similar range of previous studies, but the methodologies used to obtain them were different, as all of these studies generated discriminant function equations. Nonetheless, as each population has different genetic composition, physical activities and biomechanics, it is important to obtain equations for each population. However, this study also possess some limitations like the lack of external validation. Having another collection for testing would help reinforce the accuracy. This would be a possible next step for future projects.
Studies conducted for sex estimation using the most
ideal equation in each population were applied to different populations by different authors13,19, revealed a lower accuracy in classifying the sexes.16 Therefore, from the LR analysis, it can be suggested that patella provides an important significance in determining sex. It also signifies the importance of population specific determinator. This also provides the forensic anthropologist an alternative approach when profiling an unknown remain.
While, the logistic regression yielded a strong predictive value of 80%, the performance is lower compared
to pelvis and skull at approximately 92% to 95%. Thus, this suggest that patella while it can fulfill the role of being a supplementary bone for sex estimation, it cannot replace the primary models. Locally and internationally, the findings from studies conducted on patella has expressed it to be a viable bone in aiding the sex estimation procedure but only when the primary alternatives are missing, fragmented or unsuccessfully preserved.
Logistic regression and discriminant function analysis are the prevailing methods in sex estimation but as times have been changing and there is an integration of artificial intelligence into every aspect of our daily lives. Hence, there has also been a growing tendency to use artificial learning models such as ML and DL in the field of forensic science.12
Previously, there have been two studies that utilized a ML approach predicting the best model for sex estimation: Knecht and colleagues12 for an Italian population and Bidmos and colleagues18 for a South African population with mixed European and African descent.
Comparing this current research with the work of Bidmos and colleagues7, they achieved an overall accuracy of 89.61% using Random Forest, 89.23% using AdaBoost classifier and 90.77% using the Stacking model. The approach to the training model was different as a 5-fold cross validation training was used whereas this study used 10-fold cross validation. Using 5-fold cross validation reduced the overall accuracy for each classifier. Comparing this with the Italian population12, their overall accuracy for their ML classifiers ranged from 85% to 91% for training data and 71% to 95% for their validation sample. The best classifier for their study was the Sequential minimal optimization (SMO) which is a type of support vector machine (SVM). Furthermore, their research only utilizes three parameters in classifying sex whereas, this one uses 8 parameters, broadening the horizon of computing differences. Nonetheless, the overall accuracies are in similar ranges showcasing the agreement of using ML as a means for classifying sex. Not only this, it further strengthens the applicability of using patella in sex estimation.
Furthermore, depending on the population, various
classifiers would provide a different overall accuracy. This suggests that to study the Thai population, an ML model such as Coarse tree or SVM can be used.
Extending a step further, utilizing the images of the patella for sex estimation provides a flexible alternative to the primary choice of pelvis or the skull. ResNet 18 likely gained the highest accuracy because it is composed
of residual network that help in training at deeper layer and this allows for information to skip layers allowing the model to learn more reliably.43 However, VGG19 obtained the lowest accuracy because this is model is dense with large parameters and layers that could highly lead to overfitting on small dataset like 1,000 images.44 Finally, the MobileNetV2 is a lightweight model which could reduce flexibility in trying to read complex images.45 However, looking at the results from a different perspective, the lower accuracy value can be ascribed to the type of image used in training and testing. The anterior and posterior images used in this study as seen in Fig 3 can be described to have a cluttered or noisy background. Prior studies have reported that training and testing a convolutional network with images with cluttered background can slower the training procedure and cause models to underperform.43 Thus, this can act as a limitation not only for this study but a cautious point for future researches. The lower accuracy for our study can possibly be explained by a small dataset of 1000 images because in recent research using convolutional neural network to estimate sex from MRI images of patella, they utilized approximately 6,000 images.20 Hence, a small dataset increases the risk of overfitting and weakens the generalizability factor. An MRI image is also more homogenous with a reduced background cluttering.
Thus, these differences could also explain discrepancies
and lower result yeilded.
The selection of models like the VGG19, ResNet18 and MobileNetV2 was to study a range of model. Not specifically for forensic anthropology but these models have been used to study age and sex estimation from dental radiograph in young adults, achieving an accuracy of approximately 80%.44
Computational forensics is a current developmental content that utilizes AI to identify human remains. Deep learning analysis usually employs radiographic datasets. Studies that have utilized CT scans37 and MRI45 in measuring the patella size to determine sexual dimorphism have been able to achieve high accuracy. Furthermore, another relevant study on using DL to classify sex in relation to the patella is the radiographs of tibiofemoral joints, that obtained 90.3%.46 However, this study demonstrated a more flexible approach by using photographic imagery which is an easier, practical and much more cost-effective manner. However, with the limited sample size of 1,000 images the training and model can be developed to achieve higher accuracy. It is very important to always note the sample size, population and the methodologies utilized in each study before performing any comparison.
CONCLUSION
In conclusion, this study of sexual dimorphism of the patella within the Thai population using 250 paired patellae revealed sexual dimorphism exists in the patella as all measured parameters showed significant differences. Overall, comparing the traditional approach of logistic regression to advanced computational models such as ML and DL, it is evident that all these methods achieve an accuracy of 80% and above. These findings suggest that while logistic regression can be the basis of generating a robust sex estimation equation, ML and DL models can work to enhance the classification process. This suggests that even in the absence of the primary skeletal choice for sex determination (pelvis and skull), the patella can be used.
Data Availability Statement
The access of this data was approved through Siriraj Anatomical and Anthropological Bone Research Center (Si-AABRC) procedure. Data is available on reasonable request from Si-AABRC.
ACKNOWLEDGEMENTS
The authors of this research are thankful for Siriraj Anatomical and Anthropological Bone Research Center (Si-AABRC), Department of Anatomy, Faculty of Medicine Siriraj Hospital for allowing the authors to utilize the patellae from the bone collection. The authors are also grateful for the body donation program of Department of Anatomy, Faculty of Medicine Siriraj Hospital. This research would like to acknowledge Ms. Areeya Marie
B. Wongla for her help.
DECLARATIONS
Grants and Funding Information
None.
Conflict of Interest
None.
Registration Number of Clinical Trial
Not Applicable.
Author Contributions
Conceptualization and methodology. S.H.; Investigation, S.H.; Formal analysis, S.H. P.R., P.J.,N.C.; Visualization and writing – original draft, S.H.; Writing – review and editing, P.R., N.C., P.J., S.R., R.M., N.S.; Supervision, N.S. All authors have read and agreed to the final version of the manuscript.
Use of Artificial Intelligence
This paper utilized ChatGPT-5 (Open AI) to generate initial code drafts for R-Studio and MATLAB.
ETHICAL STATEMENT
This research was approved by the Siriraj Institutional Review Board (SiRB), Faculty of Medicine Siriraj Hospital [SIRB Protocol No.331/2568 (IRB4)].
REFERENCES
Kahana T, Hiss J. Identification of human remains: forensic radiology. J Clin Forensic Med. 1997;4(1):7–15.
de Boer HH, Blau S, Delabarde T, Hackman L. The role of forensic anthropology in disaster victim identification (DVI): recent developments and future prospects. Forensic Sci Res. 2018;4(4): 303-15.
Austin D, King RE. The Biological Profile of Unidentified Human Remains in a Forensic Context. Acad Forensic Pathol. 2016;6(3):370–90.
Krishan K, Chatterjee PM, Kanchan T, Kaur S, Baryah N, Singh RK. A review of sex estimation techniques during examination of skeletal remains in forensic anthropology casework. Forensic Sci Int. 2016;261:165.e1–8.
Ahmed AA. Estimation of sex from the lower limb measurements of Sudanese adults. Forensic Sci Int. 2013;229(1-3):169.e1–7.
Phenice TW. A Newly Developed Visual Method of Sexing the Os Pubis. Am J Phys Anthropol. 1969;30(2):297–301.
Bidmos MA, Olateju OI, Latiff S, Rahman T, Chowdhury MEH. Machine learning and discriminant function analysis in the formulation of generic models for sex prediction using patella measurements. Int J Legal Med. 2023;137(2):471–85.
Memarian A, Aghakhani K, Mehrpisheh S, Fares F. Gender determination from diagnostic factors on anteroposterior pelvic radiographs. J Chin Med Assoc. 2017;80(3):161–8.
Wangdee A, Thipdet W, Prasitwattanaseree S, Singsuwan P, Mahakkanukrauh P. Efficiency of sex determination by using external morphology of the pelvis in Thai population. BSCM. 2014;53(4):175–9.
Srinak N, Sukvitchai P. Sex estimation from patella using discriminant analysis in Central Thai population. Journal/ Journal - Canadian Society of Forensic Science. 2023;56(4):231–47.
Fox A, Wanivenhaus F, Rodeo S. The Basic Science of the Patella: Structure, Composition, and Function. J Knee Surg. 2012;25(2): 127–41.
Knecht S, Morandini P, Biehler-Gomez L, Nogueira L, Adalian P, Cattaneo C. Sex estimation from patellar measurements in a contemporary Italian population: a machine learning approach. Int J Legal Med. 2024;139(3):1371–80.
Peckmann TR, Meek S, Dilkie N, Rozendaal A. Determination of sex from the patella in a contemporary Spanish population. J Forensic Leg Med. 2016;44:84–91.
Silva RF, Franco A, Santos D, Picoli FF, Marinho DE de A. Human identification through the patella—Report of two cases. Forensic Sci Int. 2014;238:e11–4.
Chompoophuen H, Tipmala J, Duangto P, Mahakknukrauh
P. Sex Determination from the Patella in a Thai Population. Int J Morphol. 2024;42(4):891–7.
Maio C, Cunha E, Navega D. Metric analysis of the patella for sex estimation in a Portuguese sample. Forensic Sci Res. 2024; 9(2):owae015.
Nagarjuna K, Mamatha K, Venkateswarlu B. Patellar Anthropometry in Sex Differentiation - A Study in the Southern Part of Andhra Pradesh, India. Indian Journal of Forensic Medicine and Toxicology. 2021;15(2):3113–8.
Olateju OI, Philander I, Bidmos MA. Morphometric analysis of the patella and patellar ligament of South Africans of European ancestry. South African Journal of Science. 2013;109(9/10):1–6.
Bidmos MA, Steinberg N, Kuykendall KL. Patella measurements of South African whites as sex assessors. Homo. 2005;56(1):69–74.
Cavlak N, Çınarer G, Erkoç MF, Kılıç K. Sex estimation with convolutional neural networks using the patella magnetic resonance image slices. Forensic Sci Med Pathol. 2025;21(2):628–39.
Dayal MR, Bidmos MA. Discriminating Sex in South African Blacks Using Patella Dimensions. J Forensic Sci. 2005;50(6): 1294-7.
The jamovi project (2024). jamovi (Version 2.5) [Computer Software]. Available from: https://www.jamovi.org
Govindaram D, Bharanidharan R, Ramya R, Rameshkumar A, Priyadharsini N, Rajkumar K. Root Length: As a determinant tool of sexual dimorphism in an ethnic Tamil population. J Forensic Dent Sci. 2018;10(2):96-100.
Bobak CA, Barr PJ, O’Malley AJ. Estimation of an inter-rater intra-class correlation coefficient that overcomes common assumption violations in the assessment of health measurement scales. BMC Med Res Methodol. 2018;18(1):93.
Koo TK, Li MY. A Guideline of Selecting and Reporting Intraclass Correlation Coefficients for Reliability Research. J Chiropr Med. 2016;15(2):155–63.
Shrout PE, Fleiss JL. Intraclass correlations: uses in assessing rater reliability. Psychol Bull. 1979;86(2):420–8.
RStudio Team (2020). RStudio: Integrated Development for
R. RStudio, PBC, Boston, MA, USA. Available from: http:// www.rstudio.com
Verma R, Krishan K, Rani D, Kumar A, Sharma V, Shrestha R, et al. Estimation of sex in forensic examinations using logistic regression and likelihood ratios. Forensic Science International: Reports. 2020;2:100118.
Dietrichkeit Pereira JG, Fróes Lima K, Alves da Silva RH. Mandibular Measurements for Sex and Age Estimation in Brazilian Sampling. Acta Stomatologica Croatica. 2020;54(3):294–301.
The MathWorks Inc. MATLAB version: 9.14 (R2023a), Natick, Massachusetts: The MathWorks Inc., 2023. Available from: https://www.mathworks.com
Frank E, Hall MA, Witten IH. The WEKA Workbench. Online Appendix for "Data Mining: Practical Machine Learning Tools and Techniques", Morgan Kaufmann, Fourth Edition, 2016.
Sangchay N, Dzetkuličová V, Zuppello M, Chetsawang J. Consideration of Accuracy and Observational Error Analysis in Pelvic Sex Assessment: A Study in a Thai Cadaveric Human Population. Siriraj Med J. 2022;74(5):330–9.
Abdelaziz S, Khattab HM, AbdelHaq NA. Prediction of Sex
from Patellar Parameters Obtained by Magnetic Resonance Imaging in a Sample of Adult Egyptians. Zagazig Journal of Forensic Medicine and Toxicology. 2024;22(2):1–16.
Kim Y-M, Joo Y-B. Patellofemoral Osteoarthritis. Knee Surg Relat Res. 2012;24(4):193–200.
Nieves JW, Formica C, Ruffing J, Zion M, Garrett P, Lindsay R, et al. Males Have Larger Skeletal Size and Bone Mass Than Females, Despite Comparable Body Size. J Bone Miner Res. 2004;20(3): 529–35.
Introna F, Di Vella G, Campobasso CP. Sex determination by discriminant analysis of patella measurements. Forensic Sci Int. 1998;95(1):39–45.
Shang P, Zhang L, Hou Z, Bai X, Ye X, Xu Z, et al. Morphometric measurement of the patella on 3D model reconstructed from CT scan images for the southern Chinese population. Chin Med J (Engl). 2014;127(1):96–101.
Kim MH, Yoo MJ, Seo JB, Park HG, Shim SH. Statistical Analysis of the Patellar Thickness in Adults by MRI. J Korean Orthop Assoc. 2005;40(6):646-51.
Pontoh LA, Dilogo IH, Rahyussalim AJ, Widodo W, Edwin R, Fiolin J. Evaluation of patellar dimension and Bristol Index in Asian population: AnMRIstudy. AnnMedSurg(Lond). 2021;72:103072–2.
Kanchan T, Menezes RG, Moudgil R, Kaur R, Kotian MS, Garg RK. Stature estimation from foot dimensions. Forensic Sci Int. 2008;179(2-3):241.e1–5.
Bartholdy BP, Sandoval E, Hoogland MLP, Schrader SA. Getting Rid of Dichotomous Sex Estimations: Why Logistic Regression Should be Preferred Over Discriminant Function Analysis. J Forensic Sci. 2020;65(5):1685–91.
Antonogeorgos G, Panagiotakos DB, Priftis KN, Tzonou A. Logistic Regression and Linear Discriminant Analyses in Evaluating Factors Associated with Asthma Prevalence among 10- to 12-Years-Old Children: Divergence and Similarity of the Two Statistical Methods. Int J Pediatr. 2009;2009:952042.
Kc K, Yin Z, Li D, Wu Z. Impacts of Background Removal on Convolutional Neural Networks for Plant Disease Classification In-Situ. Agriculture. 2021;11(9):827–7.
Adedigba AP, Adeshina SA, Aina OE, Aibinu AM. Optimal hyperparameter selection of deep learning models for COVID-19 chest X-ray classification. Intell Based Med. 2021;5:100034.
Kalaiselvi K, Kasthuri M. Tuning VGG19 hyperparameters for improved pneumonia classification. The Scientific Temper. 2024;15(02):2231–7.
Ukwandu O, Hindy H, Ukwandu E. An evaluation of lightweight deep learning techniques in medical imaging for high precision COVID-19 diagnostics. Healthc Anal (N Y). 2022;2:100096.
Kim HE, Cosa-Linan A, Santhanam N, Jannesari M, Maros ME, Ganslandt T. Transferlearning formedical image classification: a literature review. BMC Med Imaging. 2022;22(1):69.
Yoo JH, Yi SR, Kim JH. The geometry of patella and patellar tendonmeasuredon Knee MRI. Surg Radiol Anat. 2007;29(8):623–8.
Oura P, Junno JA, Hunt DR, Lehenkari P, Tuukkanen J, Maijanen H. Deep learning insex estimation from knee radiographs – A proof-of-concept study utilizing the Terry Anatomical Collection. Leg Med (Tokyo). 2023;61:102211.