Multi-dimensional deep learning–based segmentation and volumetric assessment of sphenoid sinus fluid on postmortem CT in drowning cases

Abstract

Sphenoid sinus fluid is considered a supportive indicator of drowning in forensic medicine, but traditional manual assessment on postmortem computed tomography (PMCT) is labor-intensive and observer-dependent. Efficient, reproducible methods for quantitative evaluation are needed in forensic practice. This study developed deep learning–based approaches for the automated segmentation and volumetric estimation of sphenoid sinus fluid using PMCT images from 165 autopsy-confirmed drowning cases. Three U-Net–based models (2D, 2.5D, and 3D) were developed and evaluated against manually annotated reference standards. In the test dataset, mean Dice coefficients were 0.866 (2D), 0.869 (2.5D), and 0.798 (3D). Volumetric estimates showed no statistically significant differences from the reference standard, with strong correlations (Spearman’s ρ = 0.976–0.988). Mean absolute errors were 0.218 (2D), 0.206 (2.5D), and 0.310 ml (3D). The 2.5D approach provided the most balanced performance between segmentation accuracy and volumetric estimation. These findings demonstrate the feasibility of automated PMCT-based segmentation and volumetric quantification of sphenoid sinus fluid, enabling quantitative assessment on PMCT images prior to autopsy.

Introduction

Fluid accumulation in the sphenoid sinus is a well-known and frequently observed autopsy finding in drowning cases; however, it is considered a nonspecific finding. The diagnosis of drowning is established through a comprehensive evaluation that integrates autopsy findings, such as foam around the mouth and nose, water in the stomach, waterlogged and enlarged lungs, and diatom detection, with laboratory test results such as electrolyte and toxicological analyses, police records, and scene circumstances. Nevertheless, because these autopsy findings are nonspecific, the diagnosis of drowning remains a challenging issue in the field of forensic medicine^1,2.

Despite these limitations, various fluid-related findings observed in drowning cases have been reported to play supportive roles in the diagnosis of drowning. In particular, diatom testing and electrolyte analysis using fluid collected from the sphenoid sinus have been used as adjunct evaluation methods for suspected cases of drowning. Several studies have suggested their potential usefulness as supplementary tools for diatom detection and reference methods for differentiating between seawater and freshwater drowning^3,4.

Recently, the use of postmortem computed tomography (PMCT) in forensic medicine has continued to expand, and PMCT findings related to drowning have potential value as auxiliary tools in the diagnosis of drowning⁵. Several studies have attempted to quantitatively assess the volume and density of fluid within the paranasal sinuses on PMCT images in drowning cases⁶. In particular, for sphenoid sinus fluid, no statistically significant difference has been observed between volumes measured at autopsy and those estimated based on PMCT, and a significant correlation between the two measurements has been demonstrated, although the agreement between the two methods was relatively low. Furthermore, cases have been reported in which the sphenoid sinus fluid was not identified at autopsy but was detectable on PMCT, suggesting the potential utility of pre-autopsy PMCT for the accurate evaluation of sphenoid sinus fluid in drowning cases⁷. In such situations, quantitative evaluation of sphenoid sinus fluid on PMCT may be useful for documenting its presence and volume before autopsy and for facilitating accurate sampling for subsequent forensic examinations using sinus fluid.

However, manual segmentation and volumetric assessment of the sphenoid sinus fluid in PMCT images by observers are labor-intensive and require considerable time and effort. Automated volumetric analysis may provide a more efficient approach for evaluating sphenoid sinus fluid on PMCT. To address these limitations, artificial intelligence–based image analysis techniques have been actively introduced in the field of medical imaging, and their applications have gradually expanded to forensic medicine^8–11.

In this study, 2D, 2.5D, and 3D U-Net–based deep learning models were applied to automatically evaluate sphenoid sinus fluid in PMCT images from drowning cases. By performing automated segmentation of sphenoid sinus fluid and comparing the performance of each model, we aimed to investigate the feasibility of applying the proposed deep learning models to PMCT-based volumetric assessments in routine forensic practice.

Methods

Case selection

This study was approved by the institutional review board of the National Forensic Service (NFS), Republic of Korea (906-251029-BR-003-02). All methods were performed in accordance with the relevant guidelines and regulations, and in compliance with the Declaration of Helsinki. All data used in this study were derived from deceased individuals and were anonymized prior to analysis. Accordingly, the requirement for informed consent was waived by the institutional review board. In total, 165 drowning cases (124 males and 41 females; mean age, 56 years) were retrospectively selected from those in which both autopsy and PMCT scanning were performed at the NFS Busan Institute between August 2020 and December 2023. Cases with advanced decomposition, severe trauma, or involving infants and adolescents were excluded from the analysis (Table 1). The drowning cases occurred in various environments, including seawater (n = 122), freshwater (n = 27), bathtubs (n = 12), and swimming pools (n = 4). Drowning was diagnosed by a board-certified forensic pathologist based on a comprehensive evaluation of the autopsy findings, laboratory analysis results, police records, and circumstances at the death scene.

Table 1.

Study population and dataset composition.

Category	Value
Total number of cases	165
Sex (male/female)	124/41
Age (mean years)	56 ± 15
Training set (cases)	116
Validation set (cases)	33
Test set (cases)	16

Postmortem computed tomography scanning and reference standard segmentation

PMCT was performed using a 160-channel multidetector row CT scanner (Aquilion Prime SP; Canon Medical Systems, Otawara, Japan). All examinations were performed as whole-body scans in the supine position, and no contrast agent was administered. The scanning parameters were as follows: tube voltage, 120 kVp; tube current, 215 mAs; slice thickness, 1.00 mm, and pixel size, 0.98 mm. The acquired PMCT images were reconstructed in the axial, coronal, and sagittal planes using a soft tissue kernel. Reference standard segmentation of the sphenoid sinus fluid was performed through manual segmentation of PMCT images. Sphenoid sinus fluid was defined as a low-attenuation area distinguishable from the surrounding mucosa and bony structures that exhibited an air-fluid level. Segmentation was performed using the open-source three-dimensional (3D) analysis software 3D Slicer (version 5.8.1), and manually segmented masks were used as the reference standard for the training and performance evaluation of the deep learning–based segmentation models. Manual segmentation was performed by a single observer, and the entire segmentation process was reviewed and supervised by a forensic radiological technologist with > 10 years of experience.

Data preprocessing

The dataset used in this study consisted of 165 cases that were randomly divided into training (116 cases), validation (33 cases), and test (16 cases) sets. A total of 35,787 slices for training, 10,128 slices for validation, and 4,986 slices for testing were initially available. For the 2D and 2.5D models, slice-wise paired data comprising head PMCT images and the corresponding reference standard masks were used. Among these, only slices containing sphenoid sinus fluid were extracted and used for model development, resulting in 2,968 training slices, 918 validation slices, and 359 test slices included in the final dataset. For the 3D model, the entire head PMCT volume for each case was used without slice-level selection. All slices constituting the head PMCT volumes (35,787 for training, 10,128 for validation, and 4,986 for testing) were included to construct the 3D input data. All PMCT images and reference standard masks were converted to an 8-bit PNG format, and center cropping was applied to resize the images to a matrix size of 256 × 256. Contrast-limited adaptive histogram equalization was applied to enhance the local contrast during image preprocessing. Reference standard masks were used without preprocessing. The same preprocessing pipeline was consistently applied to the training, validation, and test datasets.

Deep learning models

To automatically segment the sphenoid sinus fluid on PMCT images, 2D, 2.5D, and 3D U-Net–based deep learning models were developed, and their performances were compared (Fig. 1). All the models were trained using the same reference standard data, and the differences in segmentation performance were evaluated by varying the dimensionality of the input data (Table 2).

Fig. 1.

Schematic overview of the 2D, 2.5D, and 3D segmentation models. The 2D model performs slice-wise segmentation using a single axial slice as input. The 2.5D model takes three consecutive axial slices as input and generates a segmentation for the center slice. The 3D model processes a volumetric CT patch and produces a volumetric segmentation output.

Table 2.

Hyperparameter of the 2D, 2.5D, and 3D deep learning models.

Category	2D model	2.5D model	3D model
Input dimension	256 × 256	256 × 256	96 × 192 × 192
Input channels	1	3	1
Input description	Single axial slice	Three consecutive slices	3D volumetric patch
Network architecture	U-Net	U-Net	3D U-Net
Encoder	ResNet-34	ResNet-34	–
Pretrained weights	ImageNet	ImageNet	None
Normalization	Batch normalization	Batch normalization	Instance normalization
Activation function	ReLU	ReLU	ReLU
Loss function	BCE + Dice loss	BCE + Dice loss	BCE + Dice loss
Optimizer	AdamW	AdamW	AdamW
Learning rate	1 × 10−4	1 × 10−4	1 × 10−4
Batch size	64	64	1
Training scheme	Slice-wise	Slice-wise	Patch-based
Patch size	–	–	96 × 192 × 192
Data augmentation	Flip, rotation	Flip, rotation	Flip, rotation

2D model

The 2D model used a two-dimensional U-Net architecture that used a single PMCT slice as input. The network consisted of an encoder–decoder structure with skip connections, and ResNet-34 was used as the encoder backbone. Batch normalization and rectified linear unit (ReLU) activation functions were applied to each convolutional block. Transfer learning was adopted using ImageNet-pretrained weights, and the number of input channels was set to one. Slice-wise paired PMCT images and reference standard masks were used for training. Data augmentation, including horizontal and vertical flipping and rotation, was performed. The batch size was set to 64, and the learning rate was set to 1 × 10⁻⁴. An AdamW optimizer was used, and a combined loss function consisting of binary cross-entropy and Dice losses was applied.

2.5D model

The 2.5D model was designed to incorporate interslice spatial contextual information using three consecutive PMCT slices: the central slice, along with the immediately preceding and following slices. These slices were concatenated along the channel dimensions, resulting in the input of three channels. The network architecture and encoder were identical to those of the 2D model using the ResNet-34–based U-Net with ImageNet-pretrained weights. Batch normalization and ReLU activation functions were applied to each convolutional block. The training settings, data augmentation, loss function, and optimization methods were identical to those of the 2D model to enable a fair comparison of the performance based solely on the input dimensionality.

3D model

The 3D model was constructed based on a 3D U-Net architecture to directly utilize volumetric information from the PMCT images. 3D patches were randomly extracted from the PMCT volumes of each case and used as input, with a patch size of 96 × 192 × 192 (Z × H × W). The number of input channels was set to one. The 3D U-Net comprised an encoder–decoder structure with skip connections, and each convolutional block included 3D convolution, instance normalization, and ReLU activation functions. Transfer learning was not applied, and the model was trained using random initialization. Data augmentation, including left–right and anterior–posterior flipping and rotation, was applied. The batch size was set to one, and the learning rate was set to 1 × 10⁻⁴. The AdamW optimizer, combined binary cross-entropy, and Dice loss function were used.

Performance evaluation and statistical analyses

Performance evaluation was conducted on an independent test dataset using the best-performing model selected based on the validation performance. Inference was performed on the test dataset using the trained models according to their respective input configurations.

Segmentation performance was quantitatively evaluated using the Dice coefficient and intersection over union (IoU). The boundary accuracy was assessed by calculating the 95th percentile Hausdorff distance (HD95). Segmentation performance was evaluated on slices in which sphenoid sinus fluid was present in either the prediction or reference standard. To assess the volumetric differences between the predicted and reference standard segmentations, voxel-based volume differences were calculated and converted to milliliters (ml) using the voxel spacing information extracted from the Digital Imaging and Communications in Medicine images. The case-wise mean absolute error (MAE) was used to identify the volume differences. All evaluation metrics were calculated on a per-case basis, and the final performance was reported as the mean across cases in the test dataset:

1

where denotes the set of voxels in the predicted segmentation and denotes the set of voxels in the reference standard segmentation:

2

where denotes the set of voxels in the predicted segmentation and denotes the set of voxels in the reference standard segmentation.

3

where and represent the sets of boundary voxels of the predicted segmentation and the reference standard segmentation, respectively, and denotes the Euclidean distance between voxels and . The HD95 was calculated as the 95th percentile of the Hausdorff distance to reduce sensitivity to extreme outliers:

4

where is the total number of subjects, is the reference standard volume of sphenoid sinus fluid for the i-th subject, and is the corresponding volume estimated from the automatic segmentation.

Normality was assessed using the Shapiro–Wilk test. Statistical significance between the reference standard and predicted volumes was evaluated using the Wilcoxon signed-rank test, and correlations between volumetric measurements were assessed using Spearman’s rank correlation analysis. The agreement between the predicted and reference standard volumes was further evaluated using Bland–Altman analysis. Statistical significance was defined as p < 0.05. All statistical analyses were performed using OriginPro 2025(OriginLab Corporation, USA).

Results

Study population and dataset characteristics

In total, 165 drowning cases were included in this study and randomly divided into training, validation, and test datasets. The test dataset consisted of 16 cases, and volumetric analyses were performed at the case-based level. The 2D and 2.5D models were evaluated using slice-wise paired PMCT images and the corresponding reference standard masks, whereas the 3D model used the entire head PMCT volume of each case as the input. All performance evaluations were conducted on an independent test dataset that was not used during the model training.

Segmentation performance of 2D, 2.5D, and 3D models

On the test dataset, the 2D model achieved a mean Dice coefficient of 0.866 ± 0.072 and a mean IoU of 0.796 ± 0.097, with a mean HD95 of 2.322 ± 3.679 mm. The 2.5D model showed a mean Dice coefficient of 0.869 ± 0.072, a mean IoU of 0.800 ± 0.096, and a mean HD95 of 2.301 ± 3.517 mm. In contrast, the 3D model demonstrated a mean Dice coefficient of 0.798 ± 0.119, a mean IoU of 0.740 ± 0.133, and a mean HD95 of 14.135 ± 44.375 mm (Table 3).

Table 3.

Segmentation performance of the 2D, 2.5D, and 3D models on the test dataset.

Model	Dice coefficient	IoU	Mean Hausdorff95 (mm)
2D model	0.866 ± 0.072	0.796 ± 0.097	2.322 ± 3.679
2.5D model	0.869 ± 0.072	0.800 ± 0.096	2.301 ± 3.517
3D model	0.798 ± 0.119	0.740 ± 0.133	14.135 ± 44.375

Estimation accuracy and agreement between predicted and reference standard volumes

A summary of the reference standard sphenoid sinus fluid volumes in the 16-case test dataset is presented in Table 4. The median reference standard volume was 2.894 ml (interquartile range [IQR], 1.089–6.743 ml). The median predicted volumes were 2.994 ml (IQR, 1.427–6.809 ml) for 2D model, 2.998 ml (IQR, 1.358–6.872 ml) for the 2.5D model, and 3.045 ml (IQR, 1.316–7.020 ml) for the 3D model.

Table 4.

Characteristics of the test dataset.

Statistic	Value
Number of cases	16
Reference standard volume, median (IQR)	2.894 ml (1.089–6.743 ml)
Reference standard volume, mean ± SD	4.014 ± 3.716 ml
Reference standard volume, range (min–max)	0.607–14.391 ml

The MAEs of the volumetric estimations were 0.218 ml, 0.206 ml, and 0.310 ml for the 2D, 2.5D, and 3D models, respectively. The Wilcoxon signed-rank tests revealed no statistically significant differences between the predicted and reference standard volumes for any model (2D, p = 0.562; 2.5D, p = 0.495; 3D, p = 0.900).

The Spearman rank correlation coefficients between the predicted and reference standard volumes were 0.985, 0.988, and 0.976 for the 2D, 2.5D, and 3D models, respectively, indicating significant correlations for all models. Bland–Altman analysis showed mean differences (bias) of − 0.087 ml, − 0.064 ml, and − 0.022 ml for the 2D, 2.5D, and 3D models, respectively. The 95% limits of agreement ranged from − 0.694 to 0.521 ml, − 0.621 to 0.493 ml, and − 0.887 to 0.843 ml for the 2D, 2.5D, and 3D models, respectively (Table 5; Figs. 2, 3 and 4).

Table 5.

Volume estimation accuracy and agreement with reference standard volumes.

Model	Median predicted volume (ml)	MAE (ml)	Wilcoxon p-value	Spearman ρ
Reference standard	2.894 (IQR, 1.089–6.743)	–	–	–
2D model	2.994 (IQR, 1.427–6.809)	0.218	0.562	0.985
2.5D model	2.998 (IQR, 1.358–6.872)	0.206	0.495	0.988
3D model	3.045 (IQR, 1.316–7.020)	0.310	0.900	0.976

Fig. 2.

Distribution of volume differences between the predicted and reference standard of sphenoid sinus fluid volumes for the 2D, 2.5D, and 3D models. Volume difference was calculated as the reference standard volume minus the predicted volume for each subject.

Fig. 3.

Scatterplots showing the relationship between the predicted and reference standard sphenoid sinus fluid volumes for the 2D, 2.5D, and 3D models.

Fig. 4.

Bland–Altman plots illustrating the agreement between the predicted and reference standard of sphenoid sinus fluid volumes for the 2D, 2.5D, and 3D models.

Discussion

Fluid accumulation within the paranasal sinuses is considered a typical autopsy finding in drowning cases; however, it is widely recognized as a nonspecific finding. Nevertheless, previous studies have suggested that paranasal sinus fluid–related autopsy findings, laboratory analysis results, and PMCT-based imaging analyses may provide supportive information for the diagnosis of drowning^12–14. In particular, PMCT-based estimation of the sphenoid sinus fluid volume shows a significant correlation with autopsy-based measurements. However, most of these approaches rely on observer-dependent manual segmentation, which is inherently time-consuming and subject to interobserver variability⁷. To address these limitations, the present study applied a deep learning–based automatic segmentation approach to enable a more objective and efficient quantification of sphenoid sinus fluid on PMCT images from drowning cases.

In the comparative performance analysis, the 2D and 2.5D models demonstrated stable segmentation performances on the test dataset. The mean Dice coefficients were 0.866 and 0.869, respectively, with corresponding mean IoU values of 0.796 and 0.800, respectively. The HD95 values were also comparable between the two models (2.322 mm for the 2D model and 2.301 mm for the 2.5D model), indicating a consistent spatial accuracy of the segmentation boundaries. In contrast, the 3D model exhibited relatively lower segmentation accuracy, with a mean Dice of 0.798 and a mean IoU of 0.740, accompanied by a markedly increased mean HD95 of 14.135 mm. Notably, the large standard deviation observed for HD95 suggests the presence of outlier cases, in which the predicted boundaries substantially deviated from the reference standard, indicating considerable variability in the segmentation performance of the 3D model across cases.

These findings are related to the structural characteristics of the sphenoid sinus fluid segmentation task. A sphenoid typically consists of two asymmetrical cavities separated by a septum and exhibits a wide range of anatomical structures and variations¹⁵. In addition, the sphenoid sinus fluid represents a small target structure that is present in only a limited number of slices within the entire head PMCT volume, with the majority of slices consisting of a non-informative background. Consequently, the voxel-level class imbalance between the target region and background is exacerbated during 3D model training, which may have contributed to boundary prediction failures and increased HD95 values in certain cases.

Furthermore, unlike clinical head CT images acquired in standardized positions, the PMCT images used in this study were frequently affected by postmortem changes, such as rigor mortis, resulting in non-true lateral head positioning and distortions in craniofacial symmetry¹⁶. These postmortem imaging characteristics may increase the apparent anatomical variability compared with clinical CT images and introduce additional uncertainty during the training and inference processes of automated segmentation models. This effect may be particularly pronounced in 3D models, which rely on spatial continuity in 3D, as nonstandard positioning and deformation can compromise segmentation stability. These findings highlight the importance of constructing training datasets and preprocessing pipelines that adequately account for the specific characteristics of postmortem CT imaging.

The representative cases presented in Figs. 5 and 6 illustrate these trends. In cases with a high segmentation agreement, all models produced stable segmentation results, although minor discrepancies were observed in some slices near the adjacent paranasal sinuses or inferior boundary regions. In contrast, cases with relatively small volumes of sphenoid sinus fluid showed an increased tendency toward segmentation errors across all models. This tendency was particularly evident in slice-based approaches, such as the 2D model, in which adjacent bony structures or soft tissues are incorrectly segmented as fluids. These error patterns are likely related to the anatomical complexity of the sphenoid sinus and its proximity to the surrounding structures.

Fig. 5.

Representative examples of segmentation results from the same case and identical axial slice positions, obtained using the 2D, 2.5D, and 3D models. The images are shown in the same slice order to enable direct visual comparison across models. Red regions indicate the predicted segmentation, yellow regions indicate the reference standard segmentation, and green regions represent the overlapping areas between the prediction and reference standard.

Fig. 6.

Representative example of segmentation results showing lower agreement with the reference standard across the 2D, 2.5D, and 3D models. These cases illustrate typical segmentation errors observed in anatomically complex regions of the sphenoid sinus.

With respect to volumetric estimation performance, sphenoid sinus fluid volumes derived from deep learning–based automatic segmentation showed no statistically significant differences from the reference standard volumes across all models and demonstrated strong correlations (Spearman ρ = 0.976–0.988). The mean MAEs were 0.218 ml, 0.206 ml, and 0.310 ml for the 2D, 2.5D, and 3D models, respectively. Bland–Altman analysis further revealed that the mean differences were close to zero for all models, with the 95% limits of agreement remaining within ± 1 ml, confirming the reliability of automated volumetric estimation. These findings suggest that automatic quantification of sphenoid sinus fluid volume on PMCT images is applicable in forensic practice.

Previous studies have primarily relied on observer-dependent manual segmentation to evaluate paranasal sinus fluid^6,7, an approach that is time-consuming, labor-intensive, and vulnerable to interobserver variability. In contrast, the present study proposes an automated segmentation-based volumetric analysis framework that addresses these limitations and enables efficient quantitative assessment of sphenoid sinus fluid both in large-scale drowning case analyses and at the pre-autopsy stage.

As medical CT data inherently possess 3D characteristics, the need for volumetric 3D segmentation has been increasingly emphasized, and the application of deep learning–based analysis has gained attention in the field of forensic radiology¹⁷. In this study, the 2.5D model demonstrated the most stable performance in terms of balancing segmentation accuracy and volumetric estimation accuracy. The 2.5D approach preserves the computational efficiency of slice-based 2D models while incorporating limited contextual information from adjacent slices, thereby improving the performance with lower computational demands compared with fully 3D models¹⁸. In contrast, 3D models require substantial graphics processing unit memory and sliding-window–based inference, which may impose constraints on the computational efficiency and workflow feasibility in forensic practice involving large-scale cases and image analysis. In this context, the 2.5D approach represents a practical and meaningful alternative that bridges the theoretical advantages of 3D models and the computational efficiency of 2D models for real-world forensic applications.

This study has some limitations. First, it was conducted using data from a single institution, and the size of the test dataset was relatively limited. Second, manual segmentation of the sphenoid sinus fluid was performed by a single observer, which precluded the quantitative assessment of interobserver variability. Although all segmentation results were reviewed and supervised by a forensic radiological technologist with > 10 years of experience. Future studies incorporating multiple observers are warranted to further evaluate the reproducibility and reliability of the proposed automated segmentation model. Third, for the 2D and 2.5D models, only slices containing sphenoid sinus fluid were used for training and evaluation to mitigate class imbalance caused by the large number of negative slices in whole-head PMCT images. Therefore, the models were designed to focus on segmentation rather than automatic detection on whole-head PMCT images. Additionally, this study focused exclusively on drowning cases and did not include comparisons with non-drowning cases. Further multicenter studies with more diverse case populations are required to assess the generalizability of the proposed approach and clarify its potential applications in forensic practice. Although this study focused on the sphenoid sinus, the proposed automated segmentation approach may also be extended to other paranasal sinuses, such as the maxillary sinus, in future studies.

Conclusion

This study demonstrated the feasibility of deep learning–based automatic segmentation and quantitative volumetric estimation of sphenoid sinus fluid on PMCT images of drowning cases. The automatic volumetric estimates obtained using the 2D, 2.5D, and 3D U-Net models showed stable agreement with the reference standard. Among the evaluated approaches, the 2.5D model exhibited the most stable balance between the segmentation accuracy and volumetric estimation accuracy. By reducing observer dependency and improving analytical efficiency, automated volumetric analysis of sphenoid sinus fluid has the potential to provide PMCT-based quantitative information at the pre-autopsy stage and support large-scale analyses of drowning cases.

Author contributions

Jin-Haeng Heo: conceptualization and writing—original draft preparation. Min-Jae Kim, Seon Jung Jang and Junghye Lee: data analysis. Sang-Beom Im: data curation. Sookyoung Lee and Joo-Young Na: writing—review and editing. Yeji Kim: model development. Yongsu Yoon and Jeong-hwa Kwon: supervision and project administration.

Funding

This work was supported by the National Forensic Service (NFS2025MED01), Ministry of the Interior and Safety, Republic of Korea.

Data availability

The datasets generated and/or analyzed during the current study are available from the corresponding author upon reasonable request, except for postmortem CT imaging data, which cannot be publicly shared owing to legal and ethical restrictions.

Declarations

Competing interests

The authors declare no competing interests.