Classification of rib fracture types from postmortem computed tomography images using deep learning.

Human or time resources can sometimes fall short in medical image diagnostics, and analyzing images in full detail can be a challenging task. With recent advances in artificial intelligence, an increasing number of systems have been developed to assist clinicians in their work. In this study, the objective was to train a model that can distinguish between various fracture types on different levels of hierarchical taxonomy and detect them on 2D-image representations of volumetric postmortem computed tomography (PMCT) data. We used a deep learning model based on the ResNet50 architecture that was pretrained on ImageNet data, and we used transfer learning to fine-tune it to our specific task. We trained our model to distinguish between "displaced," "nondisplaced," "ad latus," "ad longitudinem cum contractione," and "ad longitudinem cum distractione" fractures. Radiographs with no fractures were correctly predicted in 95-99% of cases. Nondisplaced fractures were correctly predicted in 80-86% of cases. Displaced fractures of the "ad latus" type were correctly predicted in 17-18% of cases. The other two displaced types of fractures, "ad longitudinem cum contractione" and "ad longitudinem cum distractione," were correctly predicted in 70-75% and 64-75% of cases, respectively. The model achieved the best performance when the level of hierarchical taxonomy was high, while it had more difficulties when the level of hierarchical taxonomy was lower. Overall, deep learning techniques constitute a reliable solution for forensic pathologists and medical practitioners seeking to reduce workload.

Automated Detection, Segmentation, and Classification of Pericardial Effusions on Chest CT Using a Deep Convolutional Neural Network.

Pericardial effusions (PEFs) are often missed on Computed Tomography (CT), which particularly affects the outcome of patients presenting with hemodynamic compromise. An automatic PEF detection, segmentation, and classification tool would expedite and improve CT based PEF diagnosis; 258 CTs with (206 with simple PEF, 52 with hemopericardium) and without PEF (each 134 with contrast, 124 non-enhanced) were identified using the radiology report (01/2016−01/2021). PEF were manually 3D-segmented. A deep convolutional neural network (nnU-Net) was trained on 316 cases and separately tested on the remaining 200 and 22 external post-mortem CTs. Inter-reader variability was tested on 40 CTs. PEF classification utilized the median Hounsfield unit from each prediction. The sensitivity and specificity for PEF detection was 97% (95% CI 91.48−99.38%) and 100.00% (95% CI 96.38−100.00%) and 89.74% and 83.61% for diagnosing hemopericardium (AUC 0.944, 95% CI 0.904−0.984). Model performance (Dice coefficient: 0.75 ± 0.01) was non-inferior to inter-reader (0.69 ± 0.02) and was unaffected by contrast administration nor alternative chest pathology (p > 0.05). External dataset testing yielded similar results. Our model reliably detects, segments, and classifies PEF on CT in a complex dataset, potentially serving as an alert tool whilst enhancing report quality. The model and corresponding datasets are publicly available.

Towards a New Qualitative Screening Assay for Synthetic Cannabinoids Using Metabolomics and Machine Learning.

BACKGROUND: Synthetic cannabinoids (SCs) are steadily emerging on the drug market. To remain competitive in clinical or forensic toxicology, new screening strategies including high-resolution mass spectrometry (HRMS) are required. Machine learning algorithms can detect and learn chemical signatures in complex datasets and use them as a proxy to predict new samples. We propose a new screening tool based on a SC-specific change of the metabolome and a machine learning algorithm. METHODS: Authentic human urine samples (n = 474), positive or negative for SCs, were used. These samples were measured with an untargeted metabolomics liquid chromatography (LC)-quadrupole time-of-flight-HRMS method. Progenesis QI software was used to preprocess the raw data. Following feature engineering, a random forest (RF) model was optimized in R using a 10-fold cross-validation method and a training set (n = 369). The performance of the model was assessed with a test (n = 50) and a verification (n = 55) set. RESULTS: During RF optimization, 49 features, 200 trees, and 7 variables at each branching node were determined as most predictive. The optimized model accuracy, clinical sensitivity, clinical specificity, positive predictive value, and negative predictive value were 88.1%, 83.0%, 92.7%, 91.3%, and 85.6%, respectively. The test set was predicted with an accuracy of 88.0%, and the verification set provided evidence that the model was able to detect cannabinoid-specific changes in the metabolome. CONCLUSIONS: An RF approach combined with metabolomics enables a novel screening strategy for responding effectively to the challenge of new SCs. Biomarkers identified by this approach may also be integrated in routine screening methods.

Reconstruction of full femora from partial bone fragments for anthropological analyses using statistical shape modeling.

OBJECTIVES: Due to taphonomic processes such as burial, fire, or animal activity, bones are often found incomplete, which can pose problematic for establishing the biological profile of the deceased using anthropological methods. The aim of this study is to test the feasibility of using statistical shape modeling (SSM) to reconstruct full femora from simulated partial femora and determine the accuracy of the reconstruction. Moreover, we assess the accuracy of sex estimation and the degree of stature error added based on the reconstructed femur using different anthropological methods. METHODS: A total of 42 (28 female, 14 female) 3D models of left femora extracted from computed tomography (CT) scans were used. We performed a leave-one-out cross-validation (LOOCV) where 41 bones were used to build the SSM and one bone was used for testing. This bone was cut in 1 cm steps proximally, distally and from both ends up to 10 cm, reconstructed using SSM, and tested using the methods established by Stewart and Purkait (2005), Trotter and Gleser (1952), as well as a method based on SSM. with landmarks being automatically identified. RESULTS: The error induced by reconstructing the femur to the length measurements was low, which translated into useful stature estimations (single sided cuts up to 10 cm: 0.4-1.1%, double sided<2% for cuts shorter than 6 cm). Using Purkaits method for sex estimation on reconstructed bones looked promising as well (single sided: 90.5% when compared to applying Purkaits method on the original bone, double sided 78.6% (10 cm cut) to 97.6% (1-3 cm cuts)) Using SSM for sex classification looked promising as well (single sided cut: 81-85.7%, double sided cut: 59.5-85.3%) CONCLUSION: SSM can be used to reconstruct fragmented femora. These reconstructions can be used for sex and stature estimations, at the cost of lower accuracy. Using SSM might give investigators an additional tool to gain information about the biological profile of a deceased in cases where the fragmentation of a femur does not allow for using other anthropological methods.

RiFNet: Automated rib fracture detection in postmortem computed tomography.

Imaging techniques are widely used for medical diagnostics. In some cases, a lack of medical practitioners who can manually analyze the images can lead to a bottleneck. Consequently, we developed a custom-made convolutional neural network (RiFNet = Rib Fracture Network) that can detect rib fractures in postmortem computed tomography. In a retrospective cohort study, we retrieved PMCT data from 195 postmortem cases with rib fractures from July 2017 to April 2018 from our database. The computed tomography data were prepared using a plugin in the commercial imaging software Syngo.via whereby the rib cage was unfolded on a single-in-plane image reformation. Out of the 195 cases, a total of 585 images were extracted and divided into two groups labeled "with" and "without" fractures. These two groups were subsequently divided into training, validation, and test datasets to assess the performance of RiFNet. In addition, we explored the possibility of applying transfer learning techniques on our dataset by choosing two independent noncommercial off-the-shelf convolutional neural network architectures (ResNet50 V2 and Inception V3) and compared the performances of those two with RiFNet. When using pre-trained convolutional neural networks, we achieved an F1 score of 0.64 with Inception V3 and an F1 score of 0.61 with ResNet50 V2. We obtained an average F1 score of 0.91 ± 0.04 with RiFNet. RiFNet is efficient in detecting rib fractures on postmortem computed tomography. Transfer learning techniques are not necessarily well adapted to make classifications in postmortem computed tomography.

Evaluation of the mediastinal-thoracic volume ratio on postmortem computed tomography.

OBJECTIVES: The aim of this study was to measure the mediastinal-thoracic volume ratio (CTR_VOL) on PMCT as a more accurate version of traditional CTR, in order to assess the terminal positional relationship between the heart and lungs in the different causes of death with regard to age, gender, BMI, cardiomegaly, and lung expansion. MATERIALS: Two hundred fifty consecutive postmortem cases with pre-autopsy PMCT and full forensic autopsy were retrospectively evaluated. The lungs and the mediastinum were manually segmented on the PMCT data and the correspondent volumes were estimated in situ. CTR_VOL was calculated as the ratio of the mediastinal to the thoracic volume. The volume measurements were repeated by the same rater for the evaluation of the intrarater reliability. Age, gender, body weight and height, heart weight at autopsy, and cause of death were retrieved from the autopsy reports. Presence of lung expansion was radiologically evaluated in situ. RESULTS: CTR_VOL was positively associated with age and BMI but not with gender and was higher for cardiomegaly compared to normal hearts, lower for asphyxiation-related deaths compared to cardiac deaths and intoxications, and lower for cases with lung expansion. The intrarater reliability was excellent for the calculated volumes of both lungs and mediastinum. CONCLUSION: The results of the present study support CTR_VOL as a tool to assess the relationship between the heart and lungs in situ, which differs significantly between the studied cause of death categories.

An algorithm for automatically generating gas, bone and foreign body visualizations from postmortem computed tomography data.

Post mortem computed tomography (PMCT) can aid in localizing foreign bodies, bone fractures, and gas accumulations. The visualization of these findings play an important role in the communication of radiological findings. In this article, we present an algorithm for automated visualization of gas distributions on PMCT image data of the thorax and abdomen. The algorithm uses a combination of region growing segmentation and layering of different visualization methods to automatically generate overview images that depict radiopaque foreign bodies, bones and gas distributions in one image. The presented method was tested on 955 PMCT scans of the thorax and abdomen. The algorithm managed to generate useful images for all cases, visualizing foreign bodies as well as gas distribution. The most interesting cases are presented in this article. While this type of visualization cannot replace a real radiological analysis of the image data, it can provide a quick overview for briefings and image reports.

Interpretable machine learning model to detect chemically adulterated urine samples analyzed by high resolution mass spectrometry.

OBJECTIVES: Urine sample manipulation including substitution, dilution, and chemical adulteration is a continuing challenge for workplace drug testing, abstinence control, and doping control laboratories. The simultaneous detection of sample manipulation and prohibited drugs within one single analytical measurement would be highly advantageous. Machine learning algorithms are able to learn from existing datasets and predict outcomes of new data, which are unknown to the model. METHODS: Authentic human urine samples were treated with pyridinium chlorochromate, potassium nitrite, hydrogen peroxide, iodine, sodium hypochlorite, and water as control. In total, 702 samples, measured with liquid chromatography coupled to quadrupole time-of-flight mass spectrometry, were used. After retention time alignment within Progenesis QI, an artificial neural network was trained with 500 samples, each featuring 33,448 values. The feature importance was analyzed with the local interpretable model-agnostic explanations approach. RESULTS: Following 10-fold cross-validation, the mean sensitivity, specificity, positive predictive value, and negative predictive value was 88.9, 92.0, 91.9, and 89.2%, respectively. A diverse test set (n=202) containing treated and untreated urine samples could be correctly classified with an accuracy of 95.4%. In addition, 14 important features and four potential biomarkers were extracted. CONCLUSIONS: With interpretable retention time aligned liquid chromatography high-resolution mass spectrometry data, a reliable machine learning model could be established that rapidly uncovers chemical urine manipulation. The incorporation of our model into routine clinical or forensic analysis allows simultaneous LC-MS analysis and sample integrity testing in one run, thus revolutionizing this field of drug testing.

Time- and Site-Dependent Postmortem Redistribution of Antidepressants and Neuroleptics in Blood and Alternative Matrices.

Postmortem redistribution (PMR) leads to challenges in postmortem case interpretation. Particularly antidepressants and neuroleptics are expected to undergo PMR based on their physico-chemical properties. For the current study, time- and site-dependent PMR of 20 antidepressants and neuroleptics were investigated in humans (authentic cases); five of which are discussed in detail (citalopram, mirtazapine, quetiapine, risperidone and venlafaxine) along with two metabolites (9-OH-risperidone and O-desmethylvenlafaxine). Blood [femoral (pB) and heart blood (HB)] and tissue biopsy samples (lung, kidney, liver, spleen, thigh muscle and adipose tissue) were collected upon admission to the institute utilizing a computed tomography-guided sample collection workflow (t1). Approximately 24 h later (t2; mean 23 ± 9.3 h), samples from the same body regions were collected manually. Liquid chromatography-tandem mass spectrometry was used for quantification. Most antidepressants and neuroleptics showed significant time-dependent concentration changes indicating the occurrence of PMR. For the first time, two phases of redistribution in pB for quetiapine were proposed (concentration decreases in the early postmortem phase, followed by concentration increases) and contrasting existing literature, both concentration increases and decreases in pB overtime were observed for risperidone and 9-OH-risperidone. Venlafaxine and its metabolite only showed minimal concentration changes, while citalopram exhibited a trend for concentration increases and mirtazapine for concentration decreases in pB overtime. Based on time-dependent tissue data, passive diffusion processes along the muscle-to-pB, liver-to-HB and lung-to-HB concentration gradients could be proposed along with bacterial degradation. Overall, no case interpretation had to be adjusted, which suggests that PMR changes of antidepressants and neuroleptics do not seem to be relevant for forensic case interpretation within the 24 h period that was investigated. However, limitations of the current study (e.g., temperature-controlled storage of the bodies) could have led to an underestimation of occurring postmortem changes, hence, interpretation of postmortem results should always be conducted with care, considering PMR phenomena and inter-individual variability.

Potential use of deep learning techniques for postmortem imaging.

The use of postmortem computed tomography in forensic medicine, in addition to conventional autopsy, is now a standard procedure in several countries. However, the large number of cases, the large amount of data, and the lack of postmortem radiology experts have pushed researchers to develop solutions that are able to automate diagnosis by applying deep learning techniques to postmortem computed tomography images. While deep learning techniques require a good understanding of image analysis and mathematical optimization, the goal of this review was to provide to the community of postmortem radiology experts the key concepts needed to assess the potential of such techniques and how they could impact their work.

Clinical forensic height measurements on injured people using a multi camera device for 3D documentation.

Documenting the existence, size, position and shape of injuries is an important part of medical forensic examinations. In the photography of an injury, the documentation is limited to an approximation of size and position of the injury based on a ruler included in the image. The documentation of injuries can be improved with photogrammetry, which allows the creation of scaled 3D models of an injury that can be used to not only document and visualize the injury but also to match the injury with an injury-causing object. In this paper, the multicamera device "Botscan" was used to perform 3D whole-body documentation and measure the positions of injuries. A major advantage of 3D whole-body documentation compared to photography is that the former can be performed at a later stage of the investigation. This makes the whole-body 3D documentation of injuries an important tool for re-examination.

Postmortem imaging findings and cause of death determination compared with autopsy: a systematic review of diagnostic test accuracy and meta-analysis.

BACKGROUND: The aim of this study was to evaluate the sensitivity of postmortem computed tomography (PMCT), postmortem magnetic resonance imaging (PMMR) and PMCT angiography (PMCTA) compared with autopsy in cases of adult death investigations. METHODS: For this systematic review and meta-analysis, Embase, PubMed, Scopus, Web of Science and Medline were searched for eligible studies in October 2016; a follow-up literature search was conducted in March 2018. Studies referring to PMCT, PMCTA and/or PMMR of more than 3 cases with subsequent autopsy were included. Data were extracted from published texts in duplicate. The extracted outcomes were categorized as follows: soft tissue and organ findings, skeletal injuries, haemorrhages, abnormal gas accumulations and causes of death. The summary measure was sensitivity, if 3 or more studies were available. To combine studies, a random effects model was used. Variability and heterogeneity within the meta-analysis was assessed. RESULTS: Of 1053 studies, 66 were eligible, encompassing a total of 4213 individuals. For soft tissue and organ findings, there was a high pooled sensitivity with PMCTA (0.91, 95% CI 0.81-0.96), without evidence for between-study variability (Cochrane's Q test p = 0.331, I2 = 24.5%). The pooled sensitivity of PMCT+PMMR was very high in skeletal injuries (0.97, CI 0.87-0.99), without evidence for variability (p = 0.857, I2 = 0.0%). In detecting haemorrhages, the pooled sensitivity for PMCT+PMMR was the highest (0.88, 95% CI 0.35-0.99), with strong evidence of heterogeneity (p < 0.05, I2 > 50%). Pooled sensitivity for the correct cause of death was the highest for PMCTA with 0.79 (95% CI 0.52-0.93), again with evidence of heterogeneity (p = 0.062, I2 > 50%). CONCLUSION: Distinct postmortem imaging modalities can achieve high sensitivities for detecting various findings and causes of death. This knowledge should lead to a reasoned use of each modality. Both forensic evidence and in-hospital medical quality would be enhanced.

Communicating 3D data-interactive 3D PDF documents for expert reports and scientific publications in the field of forensic medicine.

INTRODUCTION: Modern forensic investigations increasingly revert to 3D imaging techniques, such as computed tomography, magnetic resonance imaging, and 3D surface imaging. Findings are therefore often based on 3D data sets; however, this information is commonly reported and communicated within 2D imagery. The use of interactive 3D PDFs is already established in the scientific community but has yet to be implemented in the field of forensic medicine. METHODS AND MATERIALS: Three example cases were chosen to serve as exemplary data for the most commonly applied imaging techniques in postmortem imaging. 3D surface models were created from postmortem magnetic resonance imaging (PMMR), postmortem computed tomography (PMCT), and 3D surface imaging data sets. RESULTS: PMMR revealed a space-occupying subdural hemorrhage that led to ipsilateral compression of the brain tissue of the right hemisphere. PMCT displayed a defect in the skull on the left side of the temporal bone. 3D surface imaging data displayed a patterned discoloration on the inside of the left forearm. DISCUSSION: Interactive 3D PDFs offer the possibility to communicate 3D information to the reader while maintaining all the benefits of a regular 2D PDF. With Adobe Acrobat, the reader can interactively navigate through 3D data sets and create sufficient depth cues to generate a realistic 3D perception of the data. CONCLUSION: The interactive 3D PDF is a useful extension of standard 2D PDFs and has the potential to communicate 3D data to the reader in a more complete, more comprehensible, and less subjective manner than 2D PDFs.

CT-based sex estimation on human femora using statistical shape modeling.

OBJECTIVES: Estimating the sex of decomposed corpses and skeletal remains of unknown individuals is one of the first steps in the identification process in forensic contexts. Although various studies have considered the femur for sex estimation, the focus has primarily been on a specific single or a handful of measurements rather than the entire shape of the bone. In this article, we use statistical shape modeling (SSM) for sex estimation. We hypothesize that the accuracy of sex estimation will be improved by using the entire shape. MATERIALS AND METHODS: For this study, we acquired a total of 61 femora from routine postmortem CT scans at the Institute for Forensic Medicine of the University of Zurich. The femora were extracted using segmentation technique. After building a SSM, we used the linear regression and nonlinear support vector machine technique for classification. RESULTS: Using linear logistic regression and only the first principal component of the SSM, 76% of the femora were correctly classified by sex. Using the first five principal components, this value could be increased to 80%. Using nonlinear support vector machines and the first 20 principal components increased the rate of correctly classified femora to 87%. DISCUSSION: Despite some limitations, the results obtained by using SSM for sex estimation in femur were promising and confirm the findings of other studies. Sex estimation accuracy, however, is not significantly improved over single or multiple linear measurements. Further research might improve the sex determination process in forensic anthropology by using SSM.

Applying virtual reality in forensics - a virtual scene walkthrough.

A major task of forensic investigations is the documentation and interpretation of evidence to reconstruct a forensically relevant incident. To accomplish this task, a scene is documented not only with photographs but also with 3D documentation technologies. The resulting 3D data are used for 3D visualization and to perform 3D reconstructions. In this article, we present an approach for using forensic 3D data in conjunction with virtual reality to perform scene walkthroughs in the context of witness or suspect interrogations. The aim is to provide a method for scene visits showing the original scene even years after the incident. These scene walkthroughs in VR can be reproduced and allow to see through the eyes of a witness by recording their behavior and actions. These recordings allow subsequent examinations and reconstruction to support the investigation and scene understanding and can be used as evidence in court.

Using virtual reality for forensic examinations of injuries.

The ability to accurately determine injury dimensions is an essential property of forensic documentation. The current standard for injury documentation is photography using a scale to approximate the injury dimensions in the image. The technical qualities of the photograph, such as orthogonality, depth of the field and sharpness of the desired area, are vital to obtaining a correct measurement. Adequate training of the forensic staff can reduce technical errors; nonetheless, there will always be some loss of information when visualizing an injury as a three-dimensional (3D) object on a two-dimensional (2D) photograph. The shortcomings of 2D photographs can be resolved by using 3D photogrammetry, which allows 3D documentation of persons and their injuries. A series of photographs has to be acquired and processed in photogrammetric software to create a photorealistic 3D model. In a prior study, a mannequin equipped with wound tattoos of known dimensions was documented with 3D photogrammetry using a multi-camera device. On the created 3D model, the dimensions of the injuries were then measured and compared to the dimensions approximated from standard forensic photographs. The results showed that the photogrammetric measurements in 3D are more accurate than the approximations performed with standard forensic photographs. In this subsequent study, the created 3D model was visualized and surveyed in virtual reality (VR), and the results were compared to the previous study. Our goal was to establish how accurately injuries can be measured in VR compared to the standard forensic photo documentation and photogrammetric method that is used on computer screens. We found that the measurements in VR are more accurate than the approximations from forensic photo documentation, but slightly less accurate than the photogrammetric measurements performed on a computer screen in dedicated software. In conclusion, photogrammetric software and virtual reality tools can both be used to make accurate size measurements of forensics-relevant injuries. Furthermore, 3D models can be visualized in varying ways allowing a much better understanding and review of injuries, even after the injury has healed.

Application of Cinematic Rendering in Clinical Routine CT Examination of Ankle Sprains.

OBJECTIVE: Cinematic rendering technique is used to generate almost photorealistic 3D reconstructions of volumetric data. The purpose of this study was to evaluate the feasibility of cinematically rendered reconstructions in routine CT examinations of ankle sprains. CONCLUSION: Cinematic rendering technique may be primarily used to deliver visual information to patients, physicians, and virtual anatomy classes. Postprocessing requires more time than traditional methods do, which can be a hindrance in clinical work.

Comparing fist size to heart size is not a viable technique to assess cardiomegaly.

PURPOSE: Several medical textbooks state that a human heart is approximately the size of that person's fist. Stating that a heart has the size of the corpse's fist is thought to signify that the heart size is normal. We formulate two hypotheses that are tested in this article. First, we hypothesize that in cases without cardiomegaly, volumes of the hand and the heart are not significantly different. Second, we hypothesize that in cases of cardiomegaly, the heart volume statistically significantly exceeds that of the hand. MATERIALS AND METHODS: We retrospectively reviewed 130 consecutive postmortem computed tomography datasets from the BLINDED starting from 01/01/2013, covering a period of approximately 3 months. Hands and hearts were segmented and their volumes estimated. We obtained the following information from the postmortem examination reports: age, sex, body length and weight, heart weight, cardiomegaly, and cause of death. RESULTS: When exploring the correlation between mean hand volume and heart volume, only in the group of the females with cardiomegaly (N=8) could a positive, statistically significant correlation be ascertained (Pearson correlation coefficient 0.753, P=.031). DISCUSSION: In this study, we demonstrated that the commonly used idea that a heart larger than the fist of a patient suggests cardiomegaly might be incorrect. Because this perception is commonly used in autopsy reports, it might lead to avoidable errors. Until further studies confirm this hypothesis, this informal measurement should no longer be taught or used.

Comparison of forensic photo-documentation to a photogrammetric solution using the multi-camera system "Botscan".

As forensic science technologies progress, digital photography has become outdated for certain documentations that require exact measurements. Recording three-dimensional objects on a two-dimensional photograph leads to a potential loss of relevant information. Photogrammetry has been utilized to record persons, objects or crime scenes and prevents this loss. Photogrammetry enables accurate documentation and visualization of events or matching of injuries and injury-causing instruments. To reduce inaccuracies during photogrammetric recording, a multi camera device, Botscan by Botspot, can be used to record living persons in three-dimensional space (3D). The device can record a full body in a fraction of a second, which leads to a significant reduction of inaccuracies due to movement. Photogrammetric measurements were compared with measurements from forensic photographs to evaluate the applicability of this device for medical forensic documentation of injuries. For this purpose, a mannequin fitted with different types of artificial injuries was used as an example. The results showed that the photogrammetric measurements obtained using the software Agisoft PhotoScan were more accurate than the measurements from the forensic photographs.