Multifaceted Preprocessing Optimization for Post-Mortem Radiomics Analysis: A Pilot Study on Forensic Age Estimation Using Proximal Femur Radiomics

Abstract

The estimation of age at death constitutes a huge chapter in Forensic Medicine and Anthropology. In the present pilot study, a novel approach employing postmortem CT radiomics is proposed, along with an attempt at a multifaceted preprocessing optimization. A total of 60 proximal right femora were segmented automatically using 3D Slicer, and Radiomics Features were extracted in 12 different preprocessing configuration settings of the following parameters: resampling: 0.5 × 0.5 × 0.5, 1 × 1 × 1; Laplacian of Gaussian kernel size: 3, 5, 7; and bin width: 32, 64. Feature stability was tested among configurations. Following feature selection, 12 Random Forest Regression Models were developed. Most feature groups and subgroups yielded similar stability distributions, with the exception of NGTDM features. Feature selection was dominated by first order features and low pass filtered wavelet decompositions. The best performing RF model achieved R2 = 0.527 and RMSE of approximately 12.8 years, built on a downsampled configuration and an intermediate Laplacian of Gaussian scale. In overview, these preliminary results appear to be promising for the development of an extensively automated model for age at death estimation, thus capable of revolutionizing Forensic Medicine and Anthropology practices.

Supplementary Information

The online version contains supplementary material available at 10.1007/s10278-025-01714-0.

Introduction

Estimation of age at death is of major importance in the forensic examination of unknown remains, as it constitutes one of four pillars of biological profiling (sex – age – stature – population affinity), and is therefore essential for positive identification. As such, it has been studied extensively by the Forensic Anthropology community, employing a number of methodological approaches and different skeletal elements. Traditional approaches that are widely used include the morphological assessment of the pubic symphysis [1, 2] and the auricular surface of the os coxa [3], the sternal end of the 4th rib [4], and the sternal end of the clavicle [5, 6]. Despite their popularity in the field, it is well acknowledged that most existent methodologies provide extensive age ranges of several decades (e.g. [3],), which can potentially hinder rather than assist an investigation. Additionally, estimation of age in older individuals (over the age of 60 or 65) is especially challenging and unreliable [7]. More recently, similar methodologies have been developed using imaging acquisitions, both through visual, qualitative assessment of the skeletal elements and through more advanced, artificial intelligence-augmented techniques [8–14].

Postmortem imaging is a relatively new and rapidly evolving field with multiple applications within the broader scope of Forensic Medicine and Anthropology [15, 16]. A major advantage of imaging-based approaches in age estimation can be found in their non-invasive nature and their applications on both fleshed individuals and dry bone, while overriding the need for chemical or manual bone maceration [17]. At the same time, it is essential to recognize potential variations in tissue composition encountered in postmortem compared to clinical acquisitions, for example, differences in muscle attenuation possibly attributed to rigor mortis [18–20]. Several postmortem artifacts have been identified and thoroughly examined by the forensic imaging community, resulting from postmortem changes and body decomposition, most prominent in several soft tissues (e.g., gas accumulation in the subcutaneous fatty tissue) and organs (e.g., loss of cortico-medullary differentiation during cerebral autolysis) [20–23]. On the other hand, postmortem CT (PMCT) displays no motion artifacts while radiation dose reduction is not necessary [20, 23].

Radiomics analysis is a relatively new biomedical image analysis weapon in the armory of Clinical Radiology with vast potential to be transferred and employed within the field of forensics. The essence of radiomics can be summarized in the conversion of imaging acquisitions into mineable, quantitative high-dimensional data through advanced mathematical analyses and transformations [24–27]. Radiomics Features, which constitute the raw output of a radiomics pipeline, are numerical values that convey information on comprehensible concepts such as size and shape, but also on information that is otherwise imperceivable to the human eye [24, 25, 28, 29].

Even though nomenclature is still displaying variation among different extraction platforms and authors, in general, Radiomics Features can be divided into four broad categories: shape, first order, second order, and higher order features [30]. While first order features are histogram-based, second order features offer textural information through the analysis of voxel interrelationships [30]. A prime example of second order features is the GLCM (Gray Level Co-occurrence Matrix) method, which is able to characterize patterns based on the architectural arrangement of voxels along the gray level scale (https://pyradiomics.readthedocs.io/en/latest/features.html). Finally, higher order features are the result of image transformations through filter impositions (e.g., wavelets, Laplacian of Gaussian, Gabor) [30]. The significance of image preprocessing (including image discretization, interpolation, and filter imposition) is well recognized in relation to the stability and reproducibility of Radiomics Features, and recommendations on the optimization of the process are emerging [31]; however, the standardization of parameter setting warrants more research on the subject.

To date, a significant volume of radiomics research is focused on the field of Clinical Oncology to support and augment diagnosis and clinical decision-making [25, 32–37]. Information on tumor size, shape, and heterogeneity from Radiomics Features can be used to predict biological behavior as well as potential resistance to therapy [38, 39]. More recently, new radiomics-based applications are emerging in the field of forensics for purposes of postmortem interval estimation [40] and timing of injuries [41–43]. On the age estimation front, limited publications are available [44–46], employing Computed Tomography (CT) and Magnetic Resonance Imaging (MRI) acquisitions of living individuals for purposes of legal age assessment.

The utility of PMCT bone radiomics for adult age at death estimation is yet to be thoroughly explored. At the same time, despite recent scientific movements such as the Image Biomarker Standardization Initiative (IBSI) [47–50] striving towards consistency of radiomics analysis implementations, specific guidelines on image pre-processing are effectively lacking, especially for postmortem data. In fact, in most radiomics studies, feature reproducibility analysis is either not performed or performed across a very limited number of conditions, primarily focusing on the Region of Interest (ROI) segmentation step of the pipeline [51]. In the present pilot study, the value of femoral bone PMCT radiomics for age at death estimation is tested across multiple image preprocessing configuration settings. Considering the differences between postmortem and antemortem imaging and the absence of relevant published literature on the stability and utility of radiomics within the postmortem setting, this manuscript aims to provide a multifaceted study of feature reproducibility across radiomics parameters, thus setting a basis for future feature reproducibility and postmortem radiomics studies.

Materials and Methods

Dataset: Sample Selection and Imaging Acquisitions

For the purposes of this study, a total of 60 full body postmortem CT scans were retrospectively collected, and the proximal femora were selected for analysis. The selection of the femur as a region of interest (ROI) was based on the high recovery rate of the skeletal element, even in advanced decomposition scenarios. Additionally, the proximal femur–acetabulum articulation provides protection from extrinsic factors (e.g., water currents in submersion contexts), a fact that, when paired with the robustness of the bone, explains the low rate of destructive postmortem damage observed in the region.

All imaging acquisitions pertained to routine forensic casework handled by the Forensic Medicine Unit of the University of Crete between 2021 and 2024. The sample consisted of 32 adult males and 28 adult females of known age, ranging between 18 and 102 years. Mean age was similar for both sexes (66.7 years in males and 67.5 in females). The demographic composition of the sample is provided in Supplementary Material (Supplementary Table 1). Casework was initially screened to exclude juveniles (< 18 years old), cases exhibiting signs of moderate to advanced decomposition (postmortem interval between death and scanning > 48 h) and cases displaying trauma, surgical interventions or obvious pathology of the femora. According to the aforementioned excluding criteria, the right femur demonstrated higher availability in the sample and was therefore selected for further analysis. Even though the effect of laterality is not explored in the present study, structural, symmetrical and biomechanical differences between the right and left femur have been reported to be minimal or even statistically insignificant [52–56].

All imaging acquisitions were captured at the Department of Radiology and Medical Imaging of the University Hospital of Heraklion using a Siemens SOMATOM Drive 256 slice CT scanner. Slice thickness of 0.625 mm on a 512 × 512 matrix and a bone reconstruction kernel were used for all cases. All cases were anonymized prior to image acquisition, as they were part of ongoing forensic investigations, and each one of them received a unique coding number before being uploaded to the hospital server.

Radiomics Features Extraction Pipeline

Anonymized imaging acquisitions were extracted in Digital Imaging and Communications in Medicine (DICOM) format and imported in the open-source software 3D Slicer v. 5.4.0 for Windows, for segmentation of the Region of Interest (ROI) and Radiomics Features extraction. An overview of the pipeline is demonstrated in Fig. 1.

Fig. 1.

Flowchart of the radiomics features extraction pipeline, including segmentation of ROI and image preprocessing (artwork created with Ibis Paint X)

Segmentation of ROI

The right proximal femora were selected as ROI. Initially, automatic segmentation of the complete femora was performed using Total Segmentator, a fully automated deep learning segmentation algorithm available within 3D Slicer [57]. Segmentations were subsequently edited manually using the Brush, Scissors, and Eraser tools of the Segment Editor module, and manual contouring was performed if necessary. Finally, the segmented femora were cropped at the caudal-most level of the lesser trochanter to retain the proximal portion of the bone. Images were reviewed in axial, sagittal, and coronal reconstructions to ensure consistent editing of the segment and to avoid segmentation artifacts in all cases.

Image Preprocessing and Extraction of Radiomics Features

3D Slicer allows for extraction customization on three image preprocessing levels through the Slicer Radiomics module: voxel size resampling, bin width, and Laplacian of Gaussian (LoG) kernel size. For the purposes of this study, 12 configurations combining different values for the aforementioned parameters were employed for the extraction of Radiomics Features, as seen in Table 1. Although the same slice thickness was used in all imaging acquisitions of this study, both oversampling and downsampling of voxel size were performed, as no clear recommendation exists in the literature to the best of the authors’ knowledge. SitkBSpline interpolation, the default option provided in Slicer Radiomics, was applied for sample discretization. Commonly encountered bin width values (32, 64) were selected. Whereas fixing the bin number results in image normalization, fixing the bin width allows the comparison of data exhibiting different intensity value ranges [27, 31, 58]. Even though no hard evidence has been established favoring either approach, taking into account the acquisition consistency in our dataset, a fixed bin width was selected over a fixed bin number.

Table 1.

Image pre-processing configuration settings employed in the radiomics pipeline

Configuration	Resampled voxel size	Bin width	LoG kernel size
1	1 × 1 × 1	64	5
2	1 × 1 × 1	32	5
3	1 × 1 × 1	64	3
4	1 × 1 × 1	64	7
5	1 × 1 × 1	32	3
6	1 × 1 × 1	32	7
7	0.5 × 0.5 × 0.5	64	5
8	0.5 × 0.5 × 0.5	64	3
9	0.5 × 0.5 × 0.5	64	7
10	0.5 × 0.5 × 0.5	32	5
11	0.5 × 0.5 × 0.5	32	3
12	0.5 × 0.5 × 0.5	32	7

Additionally, the Laplacian of Gaussian (LoG) filtering was applied. LoG filters serve as edge enhancement tools, emphasizing shifting areas along the gray level scale, thus conveying information on the sharpness or dullness of an image [59]. Bigger LoG kernel sizes are paired with larger scale (sigma, σ) values, which allow the detection of gradual gray level scale changes over a wider radius compared to smaller sigma and kernel size values that attenuate abrupt signal intensity changes. Even though several LoG-derived features have been benchmarked by IBSI [47, 50], no guidelines have been established on kernel scaling to optimally capture the properties of different tissues. In the present endeavor, three different kernel sizes were employed to assess the effect of LoG transformations applied on a generally homogenous ROI, such as skeletal tissue. Finally, wavelet transformations were enabled in all settings, producing eight decompositions per level in all three dimensions (low and high pass filters: LLL, LLH, LHL, HLL, LHH, HLH, HHL, HHH).

Radiomics Features were subsequently extracted for each configuration through Slicer Radiomics, which encapsulates the PyRadiomics library, one of the most commonly employed feature extraction frameworks [60]. PyRadiomics is mostly IBSI-compliant, with any deviations clearly stated and explained in the library’s documentation. A total of 944 Radiomics Features were extracted for each segment in each configuration setting.

Analysis and Model Building

Radiomics Features: Feature Stability and Selection

Data was analyzed in the Python programming language. Stability of Radiomics Features was examined through the coefficient of variation (CV) metric, by calculating the standard deviation and mean of each individual feature’s numerical values for all configurations. CV was also calculated for feature groups (original and filter-based) and subgroups (first order, shape, texture, etc.) by calculating the mean of individual features’ CV. Before feature selection, all features were standardized using StandardScaler to ensure uniform scaling across variables. Regarding feature selection, significant features were selected to train each model using the SelectKBest function. The number of selected features was set to k = 6, taking into account the sample size (N = 60) and following the heuristic of approximately 10 samples per feature to avoid overfitting. The f_regression scoring function, which assesses the correlation between each feature and the target variable (age at death), was applied within SelectKBest. This univariate approach provides a straightforward and interpretable method for initial feature ranking before model training.

Model Building and Statistical Analysis

A regression task was attempted for all 12 configurations to estimate age at death. Data were normalized using the Standard Scaler function, and fivefold cross validation was performed. A total of 12 Random Forest Regression Models (RF1–RF12) were developed for each configuration setting, respectively. The selection of Random Forests was based on their capacity to capture complex, non-linear relationships despite the linear feature selection step. The regression models were implemented using scikit-learn (v1.3.0) with 100 estimators and a fixed random seed to ensure reproducibility. All other hyperparameters were left at their default values (max_depth = None, min_samples_split = 2, min_samples_leaf = 1, bootstrap = True).

To test the models’ performance, Mean Absolute Error (MAE), Mean Squared Error (MSE), Root Mean Squared Error (RMSE), and R2 (coefficient of determination) were calculated for each configuration setting. MAE was used to estimate the average deviation of the models’ predictions from the actual values, whereas MSE and RMSE were employed to identify the models’ susceptibility to outliers (larger errors). R2 was calculated to assess the proportion of variance in the target variable explained by the model.

Results

Radiomics Features Across Configuration Settings

Feature correlation heatmaps were created for all configuration settings (Fig. 2a), indicating that feature dependencies may be more affected by voxel resampling compared to other preprocessing parameters. In configurations 1–6, feature correlation appears to be increased in structured blocks (Fig. 2b), suggesting that features exhibit stronger correlations in larger groups, while oversampled configurations 7–12 exhibit more diffuse, less structured patterns of increased feature correlation (Fig. 2c).

Fig. 2.

a Heatmaps on radiomics feature correlation matrixes for configurations 1–12; darker cells corresponding to inreased feature correlation. b Heatmap for Configuration 5, demonstrating an example of a structured pattern of variation in downsampling. c Heatmap for Configuration 8, demonstrating an example of a diffuse pattern of variation in oversampling

Regarding feature stability, even though all groups display similar stability distribution, original features appear to be, as expected, more stable in comparison to LoG-based and wavelet-based features across configurations (Fig. 3a). In the present study, the most variable features of each group, excluding outliers, yielded CV < 0.3, indicating that most but not all features retain acceptable stability. Regarding feature orders, shape and first-order features display the lowest variation, while most other feature subgroups demonstrate similar variability spread (Fig. 3b). The greatest variation is exhibited in the Neighboring Gray Tone Difference Matrix (NGTDM) feature subgroup, yielding a median CV of > 0.2 and a significantly wider distribution of values.

Fig. 3.

Box plots of feature stability using the coefficient of variation (CV) metric. a Feature stability across various image filter types. b Box plots of feature subgroup stability, with NGTDM features being the least stable. GLCM, gray level co-occurrence matrix; GLDM, gray level dependence matrix; GLRLM, gray level run length matrix; GLSZM, gray level size zone matrix; NGTDM, neighbouring gray tone difference matrix

Age at Death Estimation Models: Random Forest (RF) Regression

Selected features for all RF models included original first-order features of Mean and Median and their low filtered (LLL) wavelet transformations. In all but RF9, the most significant feature was identified as the low filtered 10th percentile wavelet transformation (wavelet-LLLfirstorder10Percentile). Interestingly, LoG-based features were represented only in models built using the bigger LoG kernel sizes.

Overall, model differences as reflected in performance statistics are subtle. Taking into consideration the scientific task at hand, namely forensic age estimation, models demonstrating consistent performances (as reflected on MAE values) with the lowest tendencies for isolated, bigger prediction errors (as reflected on MSE and RMSE values) were opted for. Along those lines, models RF1 and RF2 demonstrated the highest R2 values, with RF2 slightly outperforming RF1 by achieving lower MSE and RMSE values. Models RF4 and RF6 displayed equal performance, yielding lower MAE values compared to RF1 and RF2 but higher MSE and RMSE values. This indicates that RF4 and RF6 produce smaller average errors; however, they are more prone to isolated, significant errors, which would make the model less optimal compared to RF1 and RF2 for the specific task at hand. In general, however, performance differences between the aforementioned models are minimal, with most exhibiting an RMSE of approximately 12–13 years. Oversampling (using a resampled voxel size of 0.5 × 0.5 × 0.5 mm) appears to negatively affect model performance. LoG kernel sizes of 7 appear to be the least favorable for model performance. Model performances for all configurations are demonstrated in Table 2 and Fig. 4. In all models, a trend of overestimating age for younger individuals (< 60 years) and underestimating age for older individuals (> 60 years) is observed, in concordance with results reported in other studies [44]. Interestingly, oversampled models (RF7–RF12), even though still underestimating age for elderly individuals, appear to yield smaller residual between actual and predicted age, most notable in the > 80 age group.

Table 2.

Model performance statistics for RF1–RF12. Best performing models (RF1, RF2) are highlighted (italic script)

	MAE (mean)	MAE (std)	MSE (mean)	MSE (std)	RMSE (mean)	RMSE (std)	R2 (mean)	R2 (std)
RF1	10.808	0.585	166.555	25.013	12.866	1.012	0.527	0.119
RF2	10.870	1.039	165.676	29.540	12.816	1.196	0.541	0.070
RF3	11.563	0.278	183.741	16.780	13.542	0.607	0.472	0.134
RF4	10.746	0.831	169.299	26.649	12.973	1.006	0.523	0.097
RF5	11.563	0.278	183.741	16.780	13.542	0.607	0.472	0.134
RF6	10.746	0.831	169.299	26.649	12.973	1.006	0.523	0.097
RF7	11.662	1.023	180.817	30.135	13.400	1.120	0.473	0.185
RF8	11.489	1.108	176.699	31.525	13.236	1.226	0.475	0.196
RF9	10.810	1.423	190.745	44.110	13.717	1.605	0.470	0.122
RF10	11.655	1.043	180.520	30.410	13.388	1.131	0.473	0.186
RF11	11.489	1.108	176.699	31.525	13.236	1.226	0.475	0.196
RF12	11.347	1.038	172.776	31.121	13.087	1.228	0.490	0.180

Fig. 4.

Actual versus predicted age regression plots showing similar distribution in all configuration settings. Best performing models are highlighted

Discussion

The task of estimating age at death for adults within the field of Forensic Anthropology is as important as it is challenging. Most traditional approaches provide vast age ranges of one to multiple decades [1–3, 7], and are especially precarious for older individuals, typically over the age of 60 [7]. Naturally, such estimates may negatively impact the course of a forensic investigation, thus hindering positive identification attempts. The presently proposed method employs the use of PMCT radiomics for adult age at death estimation using a reproducible, highly automated, and mostly unaffected by observer bias approach. Compared to traditionally employed classification-based approaches, Random Forest Regression achieved promising results, demonstrating the optimal parameters for postmortem radiomics analysis and providing a roadmap for studies aiming to optimize parameter settings for radiomics analysis.

All existent forensic age estimation methods are based on the premises that the skeleton undergoes (qualitatively and/or quantitatively) observable degenerative changes as a function of increasing age. The proximal femur is a skeletal site strongly associated with osteoarthritic lesions as well as osteoporotic fractures, mostly but not strictly observed in older females [61, 62]. In a longitudinal and generational study performed by Wang and colleagues [61], a clear decline in bone mineral density and hip strength index was observed in postmenopausal females. Similar observations are made by other authors, in part attributed to a decline of cortical bone mass as a function of increasing age [62, 63]. Although females appear to be more significantly affected, in reality similar trends in bone degradation are being observed in both sexes with males demonstrating a more gradual decrease in bone mineral density and cortical thickness [64–67].

Even though the femur is admittedly not the most popular skeletal element to be employed for age estimation purposes, a number of imaging-based studies on the subject are encountered in the literature (e.g., [9, 68–71]), focusing primarily on structural shifts and trends in bone mineral density fluctuations. Studies on proximal femora performed by Rattanachet and colleagues [68], Navega and colleagues [9], and Paschall and Ross [69] produce comparable results, as presented in Table 3. Compared to the aforementioned studies, all of which are based on imaging acquisitions of dry bone rather than fleshed remains, the present work proposes an approach that is notably more automatized, thus more reproducible, and also applicable in shorter PMI scenarios and fleshed cadavers. Table 3 summarizes selected traditional and novel methodologies currently used for age estimation, in comparison to the present study.

Table 3.

Summarized selected traditional and novel approaches for age estimation compared to the present study. AAD: Age at Death; LA: Living Age; CBCT: Cone Beam Computed Tomography; DXA: Dual-Energy X-Ray Absorptiometry

	Ref	Skeletal element	Task	Study type and approach	Analysis	Results and performance
Traditional approaches	[1]	Pubic symphysis	AAD	Dry bone; Qualitative (macroscopic) evaluation	Linear Regression on original sample	Classification into 1 of 6 phases corresponding to age ranges of approximately 10–50 years (age ranges increasing with increasing phase)
	[3]	Auricular surface	AAD	Dry bone; Qualitative (macroscopic) evaluation	Component scoring system, Linear Regression for composite score and age	Composite scores corresponding to 7 age stages with age ranges of 3 to approximately 50 years (age ranges increasing with increasing score)
	[4]	Sternal end of 4th rib	AAD	Dry bone; Qualitative (macroscopic) evaluation	Component scoring system, Cross-tabulation analysis	Composite scores corresponding to age ranges of up to approximately 40 years (age ranges increasing with increasing score)
Novel approaches	Present study	Proximal femur	AAD	PMCT; Quantitative; Automatic Segmentation; Radiomics analysis	Random Forest Regression	Best performing models producing MAE of approximately 10.8 years and RMSE of approximately 12.8 years
	[9]	Proximal femur	AAD	Plain radiographs of dry bone; Quantitative; Manual Segmentation; Appearance-based model through eigenvectors	Principal Component Analysis, K-Nearest Neighbor Search	MAE of 13.47 years in pooled sample, 11.08 years in male sample with RMSE of 14.06 years
	[68]	Proximal femur	AAD	Plain radiographs of dry bone; Quantitative; Sum of pixel values for 5 manually selected ROIs	Linear Regression Analysis	RMSE of 11.85 years for males, 12.56 years for females, and 12.67 years for the pooled sample
	[44]	Pelvis	LA	Clinical CT; Quantitative; Semi-automatic Segmentation; Radiomics analysis	Linear Regression Analysis	MAE of 8.025 and 8.7 years in training and testing samples respectively
	[46]	Mandibular condyles	LA	Clinical CBCT; Quantitative, Semi-automatic segmentation; Radiomics Analysis	Random Forest Classification analysis	Classification accuracy ranging between 72.22–100%
	[69]	Cranium and femur	AAD	DXA of dry bone; Quantitative; Bone Mineral Density	Loess Regression Analysis	RMSE of 13 years

The obvious practicality of employing the proximal femur for the task at hand must not be overlooked either. The femur, being one of the biggest and most robust elements of the human skeleton, is less frequently affected by postmortem disarticulation compared to smaller bones of the extremities, thus resulting in higher recovery rates. At the same time, the acetabulofemoral joint offers protection from taphonomic agents and potentially destructive postmortem artifacts. Finally, the size, shape, and anatomical position of the bone enable the acquisition of usable imaging data, simplifying the segmentation process, even if manually performed.

As is the case with any imaging-based approach, the non-destructive nature of the proposed method, combined with the eliminated need for manual or chemical maceration, greatly facilitates the forensic investigation. Another advantage of the present study is the high degree of automatization in the segmentation and Radiomics Features extraction processes, leaving little room for human error. At the same time, radiomics is perhaps one of the most feasible of the -omics family when it comes to cost and implementation, as it merely requires a workstation, while multiple open-source software resources are available for data extraction (e.g., 3D Slicer, used for the present study) and analysis. Radiomics Features are capable of capturing information expanding far beyond concepts such as bone mineral density to encapsulate data relating, among others, to shape, size, and texture.

Image preprocessing is a very important step of the radiomics pipeline. The goal of the present pilot study was to test the effect of several preprocessing parameters on feature stability and model performance. Interestingly, different patterns of feature correlation were observed, most notable between downsampled and oversampled configurations. For the downsampled configurations, the block-based structural correlation pattern can potentially be attributed to smoothing introduced through resampling. On the other hand, oversampling leading to more diffuse feature correlations can be the result of either the capturing of finer image details, or the introduction of artificial variance.

When comparing original and filter-based features in relation to their stability across configurations, original features are naturally less affected and therefore more stable than LoG-based and wavelet-based features. Among different feature subgroups (e.g., first order, shape, and texture), shape features appear to remain expectedly stable, in agreement with findings from other studies [72]. Notably, larger variation is observed in the NGTDM feature subgroup. These features essentially convey quantified information in relation to the difference between the gray level value of a voxel within the ROI and the average gray level value of the adjacent (neighboring) voxels (PyRadiomics documentation). Considering the edge-enhancing nature of LoG filters, which attenuate gradient or abrupt intensity changes, the aforementioned observation is not surprising.

In feature selection, first order features such as Mean and Median appear frequently, along with their transform-based derivatives. This can potentially be explained by the correlation observed between signal intensity (in this case demonstrated through mean and median values for the ROI) with bone mineral density, as suggested by other authors [73, 74]. Overall, similar features, in different rankings of significance, are selected. From the wavelet-based feature family, three-dimensional low pass filtered features (LLL) are exclusively extracted as significant. This suggests that noise reduction is favored over detail enhancement provided by high filtered wavelet decompositions. This is not surprising considering the need for amplification of subtle signal changes otherwise concealed within the homogenous, high intensity signal normally produced by skeletal tissue.

A low pass wavelet-transformed 10th Percentile feature (wavelet-LLLfirstorder10Percentile) appears to dominate consistently in all but one of the regression models. The original 10th Percentile feature also demonstrates varying but considerable significance across all configurations. The values of those features represent intensity levels at the lower end of the histogram intensity distribution. This is potentially significant as these essentially convey information on heterogeneity present within each individual ROI and in the entire sample set, for example as a result of age-related progressive degenerative skeletal changes. At this point, a clearer association between biological trends and feature representation cannot be safely deduced. Nonetheless, it would be safe to report that features capturing fluctuations of lower signal intensities of minimal noise are capable of capturing subtle age – related changes in bone quality.

Random Forest regression was selected as the primary predictive model due to its robustness to scaling, its ability to capture non-linear relationships, and its suitability for small datasets. Nonetheless, other machine learning approaches may exhibit different sensitivities to preprocessing or image variability. Such interactions should be systematically explored in future studies employing broader datasets. Default Random Forest hyperparameters were deliberately retained to ensure reproducibility as well as comparability across configurations and to maintain the focus on the influence of preprocessing. For the next steps, a combination of preprocessing optimization and systematic hyperparameter tuning should be attempted in order to further enhance and refine model performance and to capture other potentially significant interactions.

Overall, model performances across the different configuration settings are not significantly altered, while the observed subtle performance differences are consistent with the acceptable level of stability noted in the feature stability analysis. Configurations 1 and 2 (models RF1 and RF2) perform best (producing similar RMSE of approximately 12.8 years) and are therefore suggested based on the results of the present study. In order to put things into perspective, when dealing with an unidentified body in Forensic Medicine practice, any model producing consistently moderate average errors is generally more reliable than a model prone to isolated, larger errors, which could significantly hinder positive identification attempts. When it comes to resampling voxel size, oversampling appears to negatively affect model performance, possibly due to the introduction of noise in a minimally heterogenous tissue such as bone. At the same time, while downsampled models appear favored by an intermediate LoG kernel size, the best-performing oversampled model (RF12) is built on a bigger LoG kernel size. This can potentially be attributed to a degree of smoothing and noise cancellation introduced by the larger sigma LoG transformation.

While the limited sample size naturally raises concerns about potential overfitting, several steps were taken to mitigate this risk. Feature selection was intentionally conservative, using a univariate filter method (SelectKBest with k = 6), based on the commonly used heuristic of 10 samples per feature. Although feature selection was independent of the model, Random Forest was chosen due to its ensemble nature, bootstrapping, and random feature selection during training, which help limit model variance. The recurrence of similar features across configurations, particularly low-pass filtered first-order features, supports the hypothesis that the models captured stable and potentially biologically relevant trends.

A consistent trend of overestimating younger individuals and underestimating older ones was observed across all models. This pattern is likely due to a regression-to-the-mean effect, which commonly arises in regression tasks involving small, imbalanced datasets. As data were collected retrospectively and pertain to real forensic casework, equal representation of all ages could not be achieved. While the dataset spanned a wide chronological range (see Supplementary Table 1), its skew toward older adults likely contributed to this outcome. The influence of sample age distribution and model type on prediction accuracy warrants further investigation in more targeted future studies. At the same time, even though the limited sample size raises concerns regarding model generalization, the present pilot study aims to establish the feasibility of the approach and identify significant trends relating to the progression of biological age. Of course, these initial observations need to be further validated in the future employing larger datasets.

The lack of external validation datasets is recognized by the authors as a limitation of the present study. However, the primary aim of this work was to introduce and propose a PMCT bone radiomics standardization approach to be further explored in the future on a greater scale. The effect of biological sex is another implication to be taken into account for any age estimation method, especially considering the effect of hormonal changes attributed to menopause in older females [61] compared to male degenerative trends. This is something to be addressed in future steps, employing a larger sample size in order to further enhance results by creating sex-specific models. Another potential limitation of the study is the variability of population affinities within the sample, a factor that has been highlighted to affect the final estimate [75]. The effect of nutrition, body mass, drugs, and pharmaceutical agents, as well as biomechanical forces applied on the femur, must also be taken into consideration.

All imaging acquisitions utilized in the present study are postmortem and therefore a certain amount of deviation is to be expected compared to clinical or antemortem data, due to the effects of decomposition, a process initiating immediately after death. In order to avoid potentially significant inconsistencies, cases with a postmortem interval (time between death and scanning) of over 48 h were excluded in the present study. Considering the slow rate of postmortem bone degeneration, no significant differences are expected within the 48-h postmortem timeframe. Finally, the effect of antemortem versus early, intermediate, and late postmortem changes in relation to Radiomics Features stability and model performance is worth further exploration in the future through the conduction of prospective studies or through comparison of the present data (of fleshed, recently deceased individuals) to dry femora.

Despite the aforementioned limitations, the present study paves a promising path forward, towards a new methodological era in Forensic Medicine and Anthropology. First and foremost, the work at hand proves that PMCT bone radiomics can in fact be utilized for age at death estimation. Although the models demonstrate moderately good performance, increasing and stratifying the sample dataset is expected to further improve the outcome. Additionally, the relatively easy application and extensive automatization of the proposed method must be highlighted as it does not require special laboratory equipment and potentially expensive kits, as would be the case with different -omics based approaches. Finally, considering the rising popularity of imaging modalities within the field of Forensics, this approach demonstrates the ability to enhance data extraction while providing all necessary information for its replication.

Conclusion

The present study explores the potential of post mortem CT bone radiomics for estimation of age at death while at the same time attempting a multifaceted preprocessing optimization of the method. The results are promising for the development of an extensively automated model for age at death estimation, potentially capable of revolutionizing Forensic Medicine and Anthropology practices. The results of the present work suggest that radiomics features maintain acceptable stability throughout different configuration settings, with original features appearing to be expectedly more stable in comparison to LoG-based and wavelet-based features. Out of three levels of image pre-processing tested, oversampling appears to negatively affect model performance compared to downsampling, while an intermediate LoG kernel size appears to be most favorable. On the other hand, bin width alterations demonstrate no notable influence on model performance. Overall, the study successfully paves the path towards a new era of radiomics-based research within the field of Forensics by showcasing new hypothesis-driven avenues to be explored in the future.

Supplementary Information

Below is the link to the electronic supplementary material.

Author Contributions

All authors contributed to the study conception and design. Material preparation and data collection were performed by Despoina E. Flouri. Analysis was performed by Matthaios Triantafyllou and Michail E. Klontzas. The first draft of the manuscript was written by Despoina E. Flouri and all authors commented on previous versions of the manuscript. All authors read and approved the final manuscript. Michail E. Klontzas and Elena F. Kranioti supervised the project.

Data Availability

Data can be made available to researchers upon request.

Declarations

Ethics Approval

All postmortem imaging data used in the present study were acquired as part of an ongoing research protocol approved by the Research and Ethics Committee of University General Hospital of Heraklion, Crete, Greece (N. 1388/2016).

Competing Interests

The authors declare no competing interests.