Radiographic femoral measurements for sex and species classification in dogs and cats using machine learning

Abstract

Importance

Sex and species estimation from skeletal remains is important in veterinary forensic medicine, comparative anatomy, and zooarchaeology. Radiographic osteometry has been studied in dogs and cats, but machine-learning approaches have not been well evaluated for this purpose.

Objective

To assess the performance of machine-learning models and a multilayer perceptron for estimating sex and species from radiographic femoral measurements in dogs and cats.

Methods

This retrospective study analyzed pelvic radiographs of 280 animals (140 dogs and 140 cats; 70 males and 70 females of each species) using 9 radiographic measurements. Random forest, decision tree, logistic regression, extra trees, linear discriminant analysis, quadratic discriminant analysis, and a multilayer perceptron were evaluated. Feature importance was explored with Shapley additive explanations.

Results

For sex classification, the extra trees classifier showed the highest accuracy in both dogs (0.79) and cats (0.75). For species classification, logistic regression, quadratic discriminant analysis, and decision tree each achieved an accuracy of 0.89, whereas the multilayer perceptron reached 0.93 after 500 and 1,000 training cycles. The most influential variables were femoral length for sex classification in cats, left intercondylar fossa width for sex classification in dogs, and inter-femoral-head distance for species classification.

Conclusions and Relevance

Radiographic femoral measurements permit moderate sex classification and high species classification in dogs and cats. These findings support the potential use of machine-learning analysis of femoral radiographs in veterinary forensic medicine and related morphometric fields.

INTRODUCTION

In forensic medicine, determining individual-specific characteristics such as species, sex, age, size, breed, and lineage from the remains of the organisms is of great importance [1,2,3]. Determining these parameters is of great importance in the fields of veterinary forensic medicine and zooarcheology, even though this field has not been employed intensively yet as opposed to forensic medicine [2,3]. Sex estimation is the first and most important step in determining the biological profile because it is a precursor to other parameters, eliminates half of the population, reveals differences related to aging and growth, and determines variations related to the pedigree [3,4]. In veterinary forensic medicine, species estimation to distinguish pecilarities among species of similar size is also an essential part of solving forensic cases.

According to Rensch’s rule regarding sexual size dimorphism, it is reported there is direct correlation between sexual dimorphism and body size in male and inverse in female [5,6]. When making an assessment in studies on sex and species estimation, Rensch’s rule should be taken into consideration and a species-specific path should be followed.

Although other organs and tissues have been used in sex and species differentiation in living organisms, osteological studies have been the most preferred type of the study due to the fact that they are cheaper, not requiring qualified personnel, and easy to access [4,7]. Among the bones that make up the skeleton, pelvis, skull, and teeth bones, which show sexual dimorphism, have been used more than other bones [8,9]. However, it is claimed that long bones show more dimorphism than the skull after the pelvis in terms of sexual dimorphism [10].

In humans, almost all skeletal bones have been used in sex estimation studies [4,7,11,12,13,14,15]. On the other hand, this research area has been and are still being conducted using certain bones in animals [16,17,18]. Parameters obtained from imaging methods such as radiography, computed tomography, and magnetic resonance imaging have been frequently used in these studies [19]. Findings obtained from imaging methods are preferred in such studies because they are non-invasive and measurable.

Machine learning algorithms (MLAs) and artificial neural networks (ANNs) stand out from traditional statistical methods in the analysis of medical images with their properties of producing more accurate and reliable results and being easy to implement even though there are also limiting features of statistical analyses based on strong assumptions [4,20]. Machine learning algorithms come in many varieties, each involving a different evaluation of the data. These include Decision Tree (DT), which is based on the principle of processing inputs by continuously dividing them; Logistic Regression (LR), which classifies the relationship between parameters and output probability using the sigmoidal curve function; Random Forest (RF), which can generate multiple decision trees within the system; Extra Tree Classifier (ETC), which allows for randomly dividing nodes and using the entire dataset as a training set; Linear Discriminant Analysis (LDA), which reveals differences and relationships between classes; and Quadratic Discriminant Analysis (QDA), which assumes that each class has its own unique covariance matrix. These algorithms are frequently used in health-related studies in literature and are therefore used in our study [21,22,23,24,25]. In parallel with these developments, MLAs have frequently been used and continue to be used in humans for sex estimation from osteometric parameters [21,22,23,24,25]. In fact, these studies have been taken a step further, and studies comparing which of the MLAs gives results with higher accuracy and reliability have been carried out, and very valuable results have been obtained [4]. MLAs and ANNs have also been used in animals in the fields of identification, disease diagnosis and pathology [26,27]. To the best of our knowledge, there is a lack of studies utilizing MLAs and ANNs to estimate sex and species from imaging-based osteometric parameters in cats and dogs. Therefore, this study has focused on revealing the potential of MLAs and ANNs to estimate sex and species from the parameters obtained from radiography images of the thighbone in cats and dogs. The fact that certain dog breeds are very similar in size to cats makes it essential to distinguish species from these bones. The findings obtained will reveal a parameter to be used in the discrimination of two different species from the same family that are similar in size and will also determine which parameter gives more reliable and accurate results in determining sex within the species. In addition, veterinary forensic medicine will help in the identification of stray or dead animals, in the protection of wildlife, and in cases related to pets with high accuracy of sex and species identification.

METHODS

Study population

The study included pelvic radiographs obtained from 280 animals (140 dogs and 140 cats; 70 males and 70 females in each species). Pelvic radiographs obtained between January 2016 and November 2024 from client-owned dogs and cats presented to Batıkent Veterinary Hospital (Türkiye) were retrospectively reviewed. Inclusion criteria were complete visualization of both femora, skeletal maturity, and adequate image quality without positioning or motion artifacts. Radiographs showing fractures, severe osteoarthritis, hip luxation, congenital deformities, previous orthopedic surgery, or other conditions affecting femoral morphology were excluded. When multiple radiographs from the same animal were available, only one radiograph was included to avoid duplication. Bilateral femoral measurements were obtained; however, each radiograph represented one individual animal and was treated as a single observational unit in the statistical analyses. The images were obtained with a Fujifilm FCR PRİMA T (FUJIFILM Medical Co., Ltd., Japan) brand device.

Ethics statements

The study was approved by the Animal Experiments Local Ethics Committee of Sivas Cumhuriyet University (approval No. 65202830-050.04.04-05), which waived the requirement for additional owner informed consent because the study was retrospective and used anonymized archival radiographic data.

Study design

The Digital Imaging and Communications in Medicine format images obtained from the hospital radiology archive system were transferred to the PD-S Viewer V 1.4.0.0 (FUJIFILM Medical Co., Ltd.) radiology measurement program, and nine parameters were measured.

Measurement parameters

The parameters were the followings; the distance between the greater trochanters, the distance between the femoral heads, the widths of the right and left femoral bodies, the widths of the right and left medial-lateral condyles, the length of the thighbone, and the widths of the right and left intercondylar fossa, as displayed in Fig. 1.

Fig. 1. Radiographic femoral measurements used for model development in dogs and cats, and Shapley additive explanation rankings.

(A) Dog radiograph. (B) Cat radiograph. (C) Sex classification in cats. (D) Sex classification in dogs. (E) Species classification in dogs and cats. F0, distance between the greater trochanters; F1, distance between the femoral heads; F2, right femoral body width; F3, left femoral body width; F4, right medial-lateral condylar width; F5, left medial-lateral condylar width; F6, femoral length; F7, right intercondylar fossa width; F8, left intercondylar fossa width.

MLAs and ANNs process

The analyses were carried out using a Monster Abra A7 V12.5 model computer. Machine-learning analyses were performed using Python (version 3.9; Python Software Foundation, USA) with the scikit-learn library (version 1.1.1; scikit-learn developers, USA). Artificial neural network modeling was conducted using TensorFlow (version 2.8; Google LLC, USA). Statistical analyses were performed using Minitab (version 17; Minitab LLC, USA). Radiographic measurements were obtained using PD-S Viewer software (version 1.4.0.0; FUJIFILM Medical Co., Ltd.).

Prior to model development, the dataset was examined for completeness, and no missing values were identified in the radiographic measurements. Categorical outcome variables (sex and species) were encoded numerically for supervised classification. The dataset was randomly divided into training (80%) and test (20%) sets using stratified sampling to preserve class distribution. Feature scaling was performed using standardization for algorithms sensitive to variable magnitude. Model training and hyperparameter selection were performed within the training dataset using predefined parameter settings based on commonly used configurations in similar biomedical classification studies. Final model performance was evaluated on the held-out test dataset using accuracy, sensitivity, specificity, and F1 score.

Internal validation was performed using a held-out test dataset; external validation with an independent dataset was not conducted because such data were not available.

The data formed the input layer of MLAs, RF, DT, LR, ETC, LDA, QDA, and the output layer formed the status of being male and female in sex estimation, and the status of being cat and dog in species estimation. The following classifiers were implemented using the scikit-learn library (version 1.1.1): DT (criterion = gini, max_depth = 5, random_state = 42), RF (n_estimators = 200, criterion = gini, max_depth = 5, max_features = sqrt, random_state = 42), ETC (n_estimators = 200, criterion = gini, max_depth = 5, random_state = 42), LR (penalty = L2, tolerance = 0.001, random_state = 42), LDA (solver = svd, tolerance = 0.0001), and QDA (reg_param = 0.0, tolerance = 0.001). Multilayer perceptron classifier (MLPC) was preferred as the ANNs model. The multilayer perceptron classifier was implemented using TensorFlow with an input layer consisting of nine neurons, two hidden layers containing five and 10 neurons, respectively, and an output layer with two neurons. The model was trained using the Adam optimizer (β₁ = 0.9, β₂ = 0.999, ε = 1×10⁻⁸) with a learning rate of 0.001 and categorical cross-entropy as the loss function. Training was performed for 100, 500, and 1,000 epochs to evaluate model convergence and stability. A dropout rate of 0.05 was applied to reduce overfitting. Model initialization was controlled using a fixed random seed (random_state = 42). Model performance was evaluated on a held-out test dataset without early stopping or additional validation splitting. A validation split of 20% from the training dataset was used during model training to monitor performance across epochs. Early stopping was applied to prevent overfitting based on validation loss monitoring.

$$\text{Accuracy}=\frac{\text{TP}}{\text{TP}+\text{FN}+\text{FP}+\text{TN}}$$  (1)

$$\text{Sensitivity}=\frac{\text{TP}}{\text{TP}+\text{FN}}$$  (2)

$$\text{Specificity}=\frac{\text{TN}}{\text{TN}+\text{FP}}$$  (3)

$$\text{F1 score}=2\frac{\text{Precision}\times \text{Recall}}{\text{Precision}+\text{Recall}}$$  (4)

where TP is the true positive, TN is the true negative, FP is the false positive, and FN is the false negative.

In addition, the Shapley Additive exPlanations (SHAP) solver of the RF algorithm was used to evaluate the effect of the parameters on the result. Although the random forest was not the top-performing classifier in terms of predictive accuracy, it was selected for SHAP analysis because SHAP values are particularly efficient and theoretically well-grounded for tree-based ensemble models. Random forest also provides stable global feature-importance estimates and is widely used as a reference interpretability framework in medical machine-learning studies. Therefore, it was considered appropriate for identifying the relative contribution of radiographic measurements to classification outcomes.

Statistical analysis

The suitability of the measurement parameters for normal distribution was tested with the Anderson-Darling test. The two-sample t-test and Mann-Whitney U test were used in comparisons in terms of sex and species. A two-sided p value < 0.05 was considered statistically significant. The Minitab 17 package program was used for statistical analyses.

RESULTS

The parameters of the 140 cat thighbones were evaluated in terms of sex, finding that the width of the right femoral body showed a normal distribution, and had a width of 5.198 ± 1.445 in male and 5.668 ± 1.305 in female cats. Two-sample t-test showed a significant sex difference (p = 0.04).

Descriptive statistics for parameters not displaying a normal distribution and comparisons between sexes were given in Table 1. It was found that the distance between the femoral heads, the width of the left femoral body, the length of the thighbone, and the widths of the right and the widths of the left intercondylar fossa had a significant sex difference (p < 0.05). Cat thighbones were analyzed in terms of sex with MLAs, and the highest accuracy was found to be 0.75 with the ETC algorithm (Table 2). The ANN’s MLPC model reached an accuracy rate of 0.79 in 100 trainings, 0.71 in 500 trainings, and 0.75 in 1,000 trainings, displayed in Table 3.

Table 1. Descriptive statistics and sex comparisons for cat and dog thighbones.

Parameters		Sex	Median (minimum–maximum; mm)	p valuea
Cat
	The distance between the greater trochanters	Male	44.720 (29.270–73.900)	0.132
		Female	43.590 (21.030–85.940)
	The distance between the femoral heads	Male	64.500 (28.780–77.010)	0.011
		Female	60.410 (26.900–82.470)
	The widths of the left femoral bodies	Male	5.440 (2.800–9.840)	0.078
		Female	5.785 (2.460–9.320)
	The widths of the right medial-lateral condyles	Male	13.690 (9.060–19.220)	0.947
		Female	14.000 (5.340–24.820)
	The widths of the left medial-lateral condyles	Male	13.600 (5.440–18.790)	0.747
		Female	14.000 (6.830–21.870)
	The length of the thighbone	Male	99.590 (55.790–116.170)	0.001
		Female	88.910 (39.250–127.900)
	The widths of the right intercondylar fossa	Male	1.550 (0.640–5.440)	0.006
		Female	1.920 (0.420–7.800)
	The widths of the left intercondylar fossa	Male	1.530 (0.520–4.720)	0.020
		Female	1.740 (0.540–3.880)
Dog
	The distance between the greater trochanters	Male	66.880 (40.660–128.330)	0.038
		Female	80.280 (43.050–158.580)
	The distance between the femoral heads	Male	94.080 (53.770–167.030)	0.296
		Female	108.700 (44.080–171.370)
	The widths of the right femoral bodies	Male	7.785 (2.330–18.660)	0.001
		Female	11.535 (3.840–19.480)
	The widths of the left femoral bodies	Male	6.935 (2.330–18.410)	0.001
		Female	11.725 (4.600–18.790)
	The widths of the right medial-lateral condyles	Male	17.090 (4.770–38.070)	0.001
		Female	28.910 (11.530–43.610)
	The widths of the left medial-lateral condyles	Male	17.730 (7.770–74.970)	0.001
		Female	28.830 (12.270–49.830)
	The length of the thighbone	Male	110.260 (70.110–240.880)	0.001
		Female	147.660 (64.330–253.160)
	The widths of the right intercondylar fossa	Male	1.085 (0.160–3.200)	0.001
		Female	2.300 (0.640–7.770)
	The widths of the left intercondylar fossa	Male	0.830 (0.520–2.460)	0.001
		Female	2.300 (0.750–8.500)

The p values marked with bold indicate statistically significant.

^aMann-Whitney U test.

Table 2. Resolution of cat and dog thighbones in terms of sex with MLAs.

Algorithms		Accuracy	Specificity	Sensitivity	F1 scorea
Cat
	Logistic Regression	0.71	0.76	0.71	0.70
	Extra Tree Classifier	0.75	0.79	0.75	0.74
	Linear Discriminant Analysis	0.68	0.70	0.68	0.67
	Quadratic Discriminant Analysis	0.68	0.80	0.68	0.64
	Decision Tree	0.68	0.69	0.68	0.67
	Random Forest	0.71	0.71	0.71	0.71
Dog
	Logistic Regression	0.68	0.80	0.68	0.64
	Extra Tree Classifier	0.79	0.81	0.79	0.78
	Linear Discriminant Analysis	0.75	0.83	0.75	0.73
	Quadratic Discriminant Analysis	0.62	0.62	0.61	0.59
	Decision Tree	0.57	0.57	0.57	0.57
	Random Forest	0.71	0.72	0.71	0.71

MLA, machine learning algorithm.

^aF1 score as proportions (0–1 scale).

Table 3. Resolution of cat and dog thighbones in terms of gender with artificial neural networks.

Training count		Accuracy	Specificity	Sensitivity	F1 scorea
Cat
	100	0.79	0.79	0.79	0.78
	500	0.71	0.71	0.71	0.71
	1,000	0.75	0.76	0.75	0.75
Dog
	100	0.64	0.64	0.64	0.64
	500	0.64	0.68	0.64	0.63
	1,000	0.71	0.82	0.71	0.69

ANN, artificial neural network.

^aF1 score as proportions (0–1 scale).

The parameters of the cat thighbones were analyzed using the SHAP solver of the RF algorithm, and it was found that the length of the thighbone provided the highest contribution in terms of sex estimation (Fig. 1C).

When the parameters of 70 female and 70 male dog thighbones were evaluated in terms of sex, it was found that not all parameters showed a normal distribution. Descriptive statistics for parameters that do not show a normal distribution and comparisons between sexes are given in Table 1. It was found that all parameters except the distance between the femoral heads had a significant difference in terms of sex (p <0.05). Dog thighbones were analyzed in terms of sex with MLAs, and the highest accuracy was found to be 0.79 with the ETC algorithm (Table 2). As a result of the ANN’s MLPC model, an accuracy rate of 0.64 was obtained in 100 trainings, 0.64 in 500 trainings, and 0.71 in 1,000 trainings (Table 3).

The parameters obtained from dog thighbones were analyzed using the SHAP solver of the RF algorithm, and it was found that the widths of the left intercondylar fossa parameter provided the highest contribution in terms of sex estimation (Fig. 1D).

Parameters obtained from 140 dog and 140 cat thighbones were evaluated using MLAs in terms of species, and the highest accuracy was found to be 0.89 with LR, QDA, and DT algorithms (Table 4). As a result of the ANN’s MLPC model, an accuracy rate of 0.86 was obtained in 100 training sessions, 0.93 in 500 training sessions, and 0.93 in 1,000 training sessions (Table 5).

Table 4. Analysis of dog and cat thighbones in terms of species using MLAs.

Algorithms	Accuracy	Specificity	Sensitivity	F1 scorea
Logistic Regression	0.89	0.90	0.89	0.89
Extra Tree Classifier	0.86	0.86	0.86	0.86
Linear Discriminant Analysis	0.88	0.89	0.88	0.87
Quadratic Discriminant Analysis	0.89	0.89	0.89	0.89
Decision Tree	0.89	0.90	0.89	0.89
Random Forest	0.86	0.86	0.86	0.86

MLA, machine learning algorithm.

^aF1 score as proportions (0–1 scale).

Table 5. Analysis of dog and cat thighbones in terms of species using ANNs.

Training count	Accuracy	Specificity	Sensitivity	F1 scorea
100	0.86	0.86	0.86	0.86
500	0.93	0.93	0.93	0.93
1,000	0.93	0.93	0.93	0.93

ANN, artificial neural network.

^aF1 score as proportions (0–1 scale).

Parameters obtained from dog and cat thighbones were analyzed using the SHAP analyzer of the RF algorithm, and it was found that the distance between the femoral heads parameter provided the highest contribution in terms of species prediction (Fig. 1E).

DISCUSSION

As far as the sex estimation studies in human are concerned, long bones are seen as the second alternative to the pelvis in terms of sexual dimorphism [4]. It is a known fact that pelvic diameters, which constitute the birth canal, are always wider in females than in males in animals as well [28]. However, the ability to estimate sex and species from long bones in cases where the integrity of the pelvis is impaired or absent in both human and veterinary medicine is of great importance in forensic medicine, veterinary forensic medicine, archeology, and zooarcheology studies [4]. For this purpose, human long bones have been used in several studies to estimate sex using traditional statistical methods. Thighbone [29,30,31,32], humerus [33,34], radius [35], ulna [36,37,38], and tibia [39,40] have been used separately, as well as combined studies employing certain peculiarities these bones [41,42,43]. It is reported that the most important limitation encountered in studies using traditional statistical methods is that they are based on assumptions [4,20].

A study conducted on the condylar width of the thigh bone in the French population, reported a significant difference between males and females, giving a very high accuracy of 95.4% in determining sex [29]. The measurements obtained in the study were analyzed using statistics. Another research evaluating the width and axial length of the femoral neck, applying logistic regression and the C4.5 algorithms, reported the sex prediction with at a rate of 82.5%–85.7% [31]. Similarly, another report conducted in the Chinese population, indicated the most dimorphic point for the thighbone was the width of the distal epiphysis [32]. This literature pointed out the fact that the most dimorphic point for American whites and blacks was found to be the width of the femoral head in previous studies, leading to the conclusion that such sex estimation studies should be population-based [32]. The most important differences that distinguish our study from these literatures in humans are that animal bones have been used, employing MLAs. The problems encountered in geometric-morphometric studies, such as being based on assumptions and the large number of measurements made, make MLAs contemporary and one step ahead in such studies. The purpose of the aforementioned studies is generally to investigate whether long bones can be an alternative in cases where first-degree dimorphic bones such as the pelvis are absent or damaged. The findings obtained with traditional statistical methods concluded that the thighbone or other long bones could be an alternative in determining sex in such cases. The measurement of radiographs with the device’s program in our study nonetheless resulted in more objective results conducted with MLAs.

With the increasing use of artificial intelligence in several areas of the scientific world, studies have been conducted in which MLA is used to predict sex in humans [20,30,44,45]. One study utilized application of MLAs in two different populations, i.e., Turkish and Egyptian, stating their ability for sex discrimination in independent populations [20]. Likewise, our study utilized six MLAs and one ANN, reaching similar results in general.

In another report evaluating the trochlea of the humerus and the depth of the olecranon fossa with MLAs, it was concluded that the trochlea showed weak sex dimorphism while the depth of the olecranon fossa could distinguish sex with high accuracy (94%) [44]. Study revealed that MLAs give satisfactory results in distinguishing sex in cats (75% in ETC) and dogs (79% in ETC) but can predict the species with high accuracy (89% in LR, QDA and DT). The relatively low results in sex prediction can be attributed to the fact that no breed discrimination was made in the evaluated parameters. Especially in dogs, the variation in body size scale among breeds makes it necessary to conduct sex estimation studies on a breed basis. Moreover, the results of our study have shown that the use of parameters belonging to a certain dog breed in sex discrimination studies to be carried out from now on will increase the accuracy rate. Consequently, comparative studies should be carried out to determine dog breeds.

Sex and species estimation studies have also been carried out in animals with the help of traditional statistical methods using skeletal bones [16,18,46,47,48,49,50]. On the other hand, as to our knowledge, there has been no literature which has employed MLAs and ANNs in sex and species estimation studies in animals especially in dogs and cats. In this context, we believe that our study will be groundbreaking in its field.

An investigation on the sex-related differences in head and pelvis morphology in cats, has determined sex differences in the coronoid process, ventral iliac spine, and ischial arch [16]. The inability to find a dimorphic point other than the mandible in the head, which is traditionally known to be dimorphic, reveals the justification of studies that foresee the use of extremity bones as an alternative in the absence of the pelvis [10]. In our study, it has been shown that the thighbone can be an alternative in determining sex in cats in cases where the pelvis is absent or its integrity is impaired (75% in ETC). In addition, advanced imaging techniques, such as tomography and magnetic resonance imaging, will enable the emergence of data that will form a reference in veterinary anatomy and veterinary forensic medicine.

As a result of the statistical analysis of 20 parameters obtained from the pelvic bones of Retriever breed dogs [17], it was determined that six of these parameters were high in males and one in females. This study presents an equation/function used to evaluate pelvic parameters based on sex. This equation/function yielded highly accurate results when applied to 20 measured parameters. The results indicate that the equation/function can be used in veterinary forensic medicine. However, the use of pelvis parameters explains the high accuracy rate. Considering the conditions where the pelvis is not found or its integrity is impaired extremity bones can be an alternative.

Another research [46] conducted on German Shepherd dogs, evaluated the parameters obtained from the humerus and radius in the foreleg and the femur and tibia in the hind leg with discriminant analysis, reporting that the radius was successful in sex identification at a high rate of 82%, and the thighbone at 93%. The same study emphasized that the dimorphism in the foreleg was size-oriented while the dimorphism in the hind leg was shape-oriented. Yet, this report stated that the role of the thighbone in sex discrimination was greater than all other bones. In this context, it is obvious that trying to make this discrimination in a bone showing shape dimorphism such as the thighbone is more valuable. This makes also clear the reason why the results in study are relatively low. It is suggested that the use of MLAs in race-based studies will positively affect the results.

In a study using geometric morphometric method on the radiographs obtained from the pelvis and femur of British short-hair cats, it was concluded that the measurement points taken from males were more outward and sharper than those of females [18]. In addition, this result emphasizes that priority should be given to the pelvis in sex estimation studies. This information is similar to human studies and coincides with the conclusion that long bones can be used in sex estimation studies when the pelvis is not available or damaged as suggested by the literature [30].

The inclusion of a wide range of breeds rather than focusing on a specific one represents a noteworthy limitation that may influence the overall accuracy of sex estimation in this study. Consequently, it is essential for future research to investigate the extent to which breed-specific variations affect these osteometric outcomes and to determine how such factors might refine the predictive precision of the models.

Femoral length reflects longitudinal skeletal growth and overall body-size scaling, both of which are influenced by sexual size dimorphism regulated through endocrine and developmental mechanisms consistent with established allometric principles such as Rensch’s rule. Differences in intercondylar fossa width may be associated with variation in cruciate ligament attachment areas and joint-stabilization demands related to locomotor biomechanics and body-weight transmission. In addition, inter-femoral-head distance represents pelvic girdle width and acetabular spacing, which differ between dogs and cats due to species-specific posture, gait pattern, and functional adaptation of the hip joint [6,19,28]. These anatomical differences likely explain the strong contribution of these parameters to sex and species discrimination observed in the present study.

As a result, this study evaluated parameters from the thighbone radiographs of cats and dogs using MLAs and ANNs, and valuable data were acquired that can be used in sex and species discrimination for veterinary anatomy and veterinary forensic medicine. In addition, as far as we know, the first study using machine learning applied to radiographic osteometric parameters to distinguish sex and species in companion animals. The fact that the parameter that gives the highest result in sex discrimination is the length of the thighbone in cats, the distance between the left intercondylar fossa in dogs, and the most distinctive parameter in species discrimination is the distance between the femoral heads, which has been determined as new data not encountered in the literature. In particular, the fact that three MLA classifiers, namely LR, QDA and DT, can predict the species with a high accuracy of 89%, and ANN with 93%, has the potential to be a reference value for veterinary forensic medicine. Accumulation of such studies by changing the material, imaging method, and MLA models used in animals, as in humans, will surely reveal useful data for determining reference ranges for anatomy and forensic medicine.