AUTOPSY Project: Forensic and Automotive Innovations

• The forensic AUTOPSY Project employs transfer learning on 2D projections from PMCT scans to achieve around 95% accuracy in objective sex estimation from skull and pelvis images.
• The automotive AUTOPSY framework implements a GDPR-driven Privacy Manager and a six-step PET-selection process to enhance privacy and data control in connected vehicle systems.
• Both initiatives illustrate tailored AI applications that stress the importance of image standardization, systematic augmentation, and contextual disambiguation for effective field deployment.

AUTOPSY Project is a name used for at least two unrelated research efforts in recent technical literature. In forensic medicine, it denotes a project that develops a fully automated, objective pipeline for determining the biological sex of cadavers from post-mortem computed tomography of the pelvis and skull, using transfer learning and standardized two-dimensional projections (Tormena et al., 21 Jun 2026). In automotive privacy engineering, it denotes a framework for supporting privacy engineering in connected and (partially) automated vehicles through a system model, a Privacy Manager, a GDPR-principle-driven privacy-enhancing-technology selection method, and a reference architecture validated with a location-based-services demonstrator (Pape et al., 22 Jul 2025). Because these projects are distinct in domain, objectives, and technical stack, the term requires contextual disambiguation.

1. Terminological scope and disambiguation

Recent literature uses “AUTOPSY” in two technically unrelated senses. One belongs to forensic radiology and skeletal profiling, where the project targets routine medicolegal sex estimation from PMCT data. The other belongs to automotive privacy engineering, where the project addresses data-flow control and PET deployment in connected vehicles. The shared label is therefore nominal rather than methodological.

Designation	Domain	Core objective
AUTOPSY Project	Forensic medicine	Automated biological sex determination from PMCT skull and pelvis images
AUTOPSY	Automotive privacy	Privacy engineering for connected and automated vehicles

A common source of confusion is to assume that “AUTOPSY” refers exclusively to autopsy automation in forensic medicine. The automotive AUTOPSY framework instead concerns GDPR-oriented privacy engineering and was developed as a technical response to the fact that compliance does not necessarily result in a privacy-friendly system (Pape et al., 22 Jul 2025). Conversely, the forensic AUTOPSY Project is not a cause-of-death system or a histopathology system; it is specifically a skeletal sex-determination pipeline built around PMCT-derived skull and pelvis morphology (Tormena et al., 21 Jun 2026).

2. Forensic AUTOPSY: problem formulation and dataset construction

In forensic medicine, the AUTOPSY Project was designed to determine the biological sex of cadavers in routine forensic workload. The motivating problem is that visual-metric anthropological analysis of sexually dimorphic bones becomes time-consuming, observer-dependent, and often unreliable in real-world casework, especially in mass disasters or in severely decomposed and traumatized remains. The project therefore uses PMCT scans of the skull and pelvis to provide rapid, reproducible sex estimates, including cases with anatomical disfigurements and imaging artifacts (Tormena et al., 21 Jun 2026).

The study analyzed 141 autopsied cadavers from the Forensic Medical Institute of Goiânia-GO, described in the technical details as 104 males and 39 females, scanned on a Siemens SOMATOM go.Now device with 0.6 mm slice thickness and no gap. Three-dimensional Volume Rendering Technique reconstructions were generated on the syngo.via workstation, and non-osseous tissues were removed via segmentation to isolate pure bone morphology. This preprocessing stage is central to the project’s definition of “objective” analysis: the network is trained on curated osseous morphology rather than on mixed soft-tissue appearance or gross scene context (Tormena et al., 21 Jun 2026).

A key methodological choice was to avoid full 3D CNNs. The project states that full 3D CNNs demand prohibitive computation and large sample sizes, so each 3D skull or pelvis was converted into a standardized set of two-dimensional frontal “profile” projections. Forensic specialists manually identified the ideal anterior “pivot” view, then extracted 11 frames spanning ±5 rotational steps, approximately ±45°, yielding 3 059 PNG images in the final repository. The curation preserved landmarks such as the subpubic angle, pubic symphysis, glabella, supraorbital margins, and mandibular angle, while excluding posterior features such as the external occipital protuberance, which lay outside the selected view. This suggests that the project treats view standardization not as a mere preprocessing convenience, but as a substantive design constraint on what anatomical dimorphism becomes learnable.

3. Model families, transfer-learning regime, and optimization strategy

The forensic AUTOPSY Project compared seven pre-trained architectures under transfer learning: YOLO26, YOLO11, ConvNeXt-Tiny, EfficientNetV2, ViT-B16, VGG16, and ResNet50. The report specifies the relevant structural characteristics of each family, including YOLO26 with a CSP-based backbone, PANet neck, and multi-scale detection heads; YOLO11 as a real-time detector with a larger backbone; ConvNeXt-Tiny as a modernized convolutional network; EfficientNetV2 as a compound-scaled architecture with squeeze-and-excitation modules; ViT-B16 as a Vision Transformer operating on $16 \times 16$ patches; VGG16 as a classic 16-layer CNN; and ResNet50 as a residual network using identity shortcuts. All networks were initialized with ImageNet-pretrained weights, or COCO in the case of YOLO, and their final classification head was adapted to the target classes (Tormena et al., 21 Jun 2026).

Training ran for up to 300 epochs with early stopping, using subject-level splits of 75% train, 10% validation, and 15% test to prevent leakage. Because the dataset was small and imbalanced, especially for the under-represented female class, the project evaluated three augmentation strategies: a custom offline Albumentations pipeline; YOLO’s native online augmentation; and a hybrid approach. The offline pipeline generated four new variants per image using RandomResizedCrop, rotations of at most $30^\circ$, RGB shifts, mild JPEG compression, Gaussian noise, horizontal flips, and sharpening, while explicitly avoiding geometric deformers that would distort bone shape. This avoidance is methodologically important because the target signal is anatomical dimorphism rather than generic texture robustness (Tormena et al., 21 Jun 2026).

Two classification scenarios were studied. The binary scenario was Male versus Female, region-agnostic. The quaternary scenario was Male Skull, Female Skull, Male Pelvis, and Female Pelvis. In the baseline comparison, modernized CNNs, specifically ConvNeXt-Tiny and EfficientNetV2, outperformed older networks, while ViT-B16 lagged, likely due to data scarcity. YOLO26 consistently topped the baseline leaderboard at approximately 91.3% image-level accuracy, with 89.4% recall, 90.5% precision, and 89.9% F1. The project then concentrated on YOLO26 Nano, Small, and Medium. Multi-criteria ranking with TOPSIS and VIKOR across 16 performance indicators identified YOLO26 Nano with default YOLO augmentations as the optimal quaternary classifier, and subsequent hyperparameter optimization with a genetic algorithm refined learning rates, momentum, regularization, and augmentation magnitudes without displacing YOLO26 Nano as the most efficient and robust solution (Tormena et al., 21 Jun 2026).

4. Performance, explainability, and limitations in the forensic pipeline

When fully trained and tested, the final quaternary YOLO26 Nano model achieved image-level results on the held-out test set of 94.89% accuracy, 95.62% precision, 92.90% recall, and 94.06% F1. At patient level, with soft voting across the 11 projections, it reached 95.65% accuracy, 97.22% precision, 92.86% recall, and 94.36% F1. A 5-fold cross-validation yielded a mean Top-1 accuracy of 90.93%, which the study interprets as evidence of stability across varied splits (Tormena et al., 21 Jun 2026).

The class-wise patient-level breakdown is notably asymmetric. Female Skull achieved 100% precision, 71.43% recall, and 83.33% F1. Male Skull achieved 88.89% precision, 100% recall, and 94.12% F1. Female Pelvis and Male Pelvis both achieved 100% precision, 100% recall, and 100% F1. Confusion matrices showed that all errors occurred within the same anatomical region, meaning female-to-male and male-to-female errors within skull or pelvis classes, but no skull/pelvis confusions. This directly counters the misconception that the model’s errors primarily arise from failure to distinguish anatomical regions. The reported failure mode is instead sex ambiguity within a correctly identified region (Tormena et al., 21 Jun 2026).

The project also tested resilience under adverse forensic conditions. Cadavers with high-energy trauma and partial bone destruction were largely classified correctly, which the paper presents as evidence for the resilience of the 2D projection approach. Explainability was addressed through Grad-CAM, which highlighted the subpubic angle and symphysis for the pelvis, and the glabella, supraorbital ridges, and mandibular angles for the skull. The study interprets these heatmaps as validation that model decisions rest on classical forensic landmarks rather than on spurious imaging cues (Tormena et al., 21 Jun 2026).

The limitations are explicit. All scans came from a single forensics institute and one CT scanner model, so generalizability across populations, scanner manufacturers, and acquisition protocols remains untested. The angular window of ±45° around the anterior view excludes posterior landmarks that could strengthen sex estimates. The paper also notes that deep learning models can still exhibit bias toward the majority class, as shown by an elevated false-negative rate for borderline female skulls. The recommended next steps are multi-center validation with diverse scanners and protocols, expanded angular coverage including lateral or posterior projections, larger and more balanced datasets including different ancestry groups, and integration into a forensic radiology software suite with audit trails, heatmap explainability, and expert-in-the-loop review. A common misunderstanding is that “objective” automatically implies bias-free performance; the study’s own discussion rejects that interpretation by foregrounding class imbalance and untested external validity (Tormena et al., 21 Jun 2026).

5. Automotive AUTOPSY: system model, Privacy Manager, and PET-selection method

The automotive AUTOPSY project addresses privacy engineering in connected and partially automated vehicles. Its point of departure is that GDPR compliance alone does not guarantee a privacy-friendly system, because, for example, user consent to process data does not itself improve privacy-friendliness. The project therefore proposes technical building blocks that improve privacy-friendliness in the automotive domain, comprising a system-of-interest model, the Privacy Manager, a PET-selection approach based on GDPR principles, an architectural framework, and a demonstrator for location-based services (Pape et al., 22 Jul 2025).

The system model formalizes the connected-vehicle ecosystem as a tuple with entities $E=\{V,U,A,S,C,T\}$, where $V$ denotes the vehicle, $U$ the user or driver, $A$ the App Manager plus applications, $S$ the Pub-Sub Service, $C$ the calculation or processing servers, and $T$ third-party storage or analytics servers. It further defines data types $D$, locations or trust domains $30^\circ$0, a trust relation $30^\circ$1, and a raw-data-flow relation $30^\circ$2. The high-level flow is Vehicle $30^\circ$3 Pub-Sub Service $30^\circ$4 Privacy Manager in HPC $30^\circ$5 App Manager plus App $30^\circ$6 optional streaming to storage servers $30^\circ$7 cloud processing. The framework states that each arrow is annotated with the relevant triples in $30^\circ$8 and protected by PETs at trust-domain boundaries (Pape et al., 22 Jul 2025).

The Privacy Manager is the core in-vehicle enforcement point. Internally it consists of a Manifest Processor, a Policy Engine, a PET Orchestrator, and an Enforcement Engine. The Manifest Processor parses each app manifest $30^\circ$9, including required data types, supported or enforced PETs, and precision or noise constraints. The Policy Engine uses $E=\{V,U,A,S,C,T\}$0 and user-consent settings $E=\{V,U,A,S,C,T\}$1 to compute privacy rules $E=\{V,U,A,S,C,T\}$2, each specifying whether raw data is allowed, which PET chain to apply, and the reuse or caching policy. The PET Orchestrator instantiates and sequences concrete PET modules such as an LDP noise-adder, ABE-encrypter, or TEE-executor, and the Enforcement Engine mediates every read or write call from the app to sensor or ECU data while logging an audit trail. The technical report associates this line of work with prior efforts by S. Pape et al. and S. Syed-Winkler et al., as cited in its references (Pape et al., 22 Jul 2025).

A distinctive feature of the automotive AUTOPSY project is its six-step PET-selection process: use-case specification and trust model; identification of relevant GDPR principles; mapping from principles to PET types; shortlisting candidate PETs; assessment of PET maturity in terms of utility, scalability, and robustness; and trade-off plus final selection. The project formalizes a mapping matrix from GDPR Article 5 principles to PET types and proposes a mixed-integer optimization that maximizes a weighted combination of utility, scalability, and robustness while ensuring that each principle is covered. This indicates that AUTOPSY is not merely an architectural proposal, but also a decision-support framework for PET choice under explicit multi-criteria constraints (Pape et al., 22 Jul 2025).

6. Automotive AUTOPSY: layered architecture and location-based-services demonstrator

The automotive architecture is organized into four layers: the Physical Layer, containing vehicle sensors, ECUs, and TCU; the Communication Layer, containing DSRC, LTE/5G, and secured channels; the Processing Layer, containing in-vehicle HPC, cloud computing, and TEEs; and the Storage Layer, containing encrypted data lakes and OEM or third-party servers. Formal interfaces are defined between apps and the Privacy Manager, between the Privacy Manager and the Pub-Sub layer, between the Privacy Manager and PET modules, and between the Privacy Manager and storage. The framework also specifies ciphertext reuse in the data lake: when $E=\{V,U,A,S,C,T\}$3 for identical $E=\{V,U,A,S,C,T\}$4 and PET, the same ciphertext can be served to multiple apps (Pape et al., 22 Jul 2025).

To validate the framework, the project implemented a location-based-services demonstrator on ARM Cortex-A53 in-vehicle hardware running Debian Linux. In the scenario, an LBS app is installed via the App Manager; its manifest declares $E=\{V,U,A,S,C,T\}$5 and support for differential privacy; and the user consents to share obfuscated location. The deployed PET is Local Differential Privacy using the Planar Isotropic Mechanism of Geng and Viswanath, with privacy budget $E=\{V,U,A,S,C,T\}$6. The Privacy Manager enforces $E=\{V,U,A,S,C,T\}$7 before releasing coordinates. Evaluation uses spatial error $E=\{V,U,A,S,C,T\}$8 and trajectory-privacy loss $E=\{V,U,A,S,C,T\}$9 (Pape et al., 22 Jul 2025).

The demonstrator reports mean spatial error of approximately 200 m at $V$0 and approximately 300 m at $V$1, with Figure 1(c) showing a graceful trade-off curve over 50 time-steps. The in-vehicle Privacy Manager introduced end-to-end latency of less than 50 ms per request, which the report states is well within LBS real-time requirements. The user interface displays requested data types, privacy risks, intended PETs, and an adjustable $V$2 slider during installation, and overlays true versus obfuscated location and red POIs during use. The project’s stated future work includes revocation through dynamic key management, quantum-safe PETs, standardizing noise-level profiles, and extending user-centric acceptance metrics into the PET-selection process (Pape et al., 22 Jul 2025).

7. Relation to broader autopsy-AI research

The forensic AUTOPSY Project sits within a larger technical movement toward computational support for medicolegal workflows. One adjacent example is FEAT, a multi-agent forensic AI system for automated cause-of-death analysis. FEAT uses a central Planner, specialized Local Solvers, a Memory and Reflection module, and a Global Solver with hierarchical retrieval-augmented generation. It fine-tunes a DeepSeek-R1-Distill-Llama-8B backbone via LoRA with rank $V$3, learning rate $V$4, cosine decay, and 10% warm-up. Its evaluation spans 7 748 autopsy records from six institutions, a curated H-RAG corpus of 6 739 expert long-form analyses, and a 1 009-case test split across six geographic cohorts. Reported gains are +3.2% over the best baseline for long-form analysis and +10.7% for short-form conclusion, with blinded validation by four senior forensic pathologists (Shen et al., 11 Aug 2025).

FEAT is not part of the AUTOPSY Project, but its section on “Lessons for an AUTOPSY Project” is notable because it recommends a Tree-of-Thought planner with adaptive depth of approximately 4, modular Local Solvers with dynamic tool routing, a low-entropy long-horizon memory and reflection loop, hierarchical RAG, domain-adaptive LoRA fine-tuning, and a lightweight human-in-the-loop layer that captures edits as feedback signals. A plausible implication is that future expansions of AUTOPSY-like forensic systems may move from single-task image classification toward multi-component forensic reasoning pipelines that preserve auditability and expert oversight (Shen et al., 11 Aug 2025).

Another adjacent development is virtual histological staining of unlabeled autopsy tissue. The RegiStain framework jointly trains a generator $V$5, discriminator $V$6, and registration network $V$7 to transform autofluorescence images into brightfield-equivalent virtual H&E while compensating for severe autolysis. The system was trained on more than 0.7 TB of paired autofluorescence-H&E data, with a final training set of 16,159 paired $V$8 patches derived from eight non-COVID pneumonia cadavers. On 100 well-stained fields of view from 10 test slides, it achieved mean SSIM of approximately 0.82 and mean PSNR of approximately 20.32 dB. The workflow reduces conventional staining delays by replacing hours-to-days of fixation and approximately 1–2 days of lab work with autofluorescence scanning of approximately 15 minutes per slide and inference of less than 5 minutes per whole-slide image on GPU (Li et al., 2023).

Taken together, these neighboring systems indicate that contemporary autopsy-related AI is expanding along at least three axes: skeletal profiling from PMCT, language-mediated cause-of-death reasoning, and virtual histological reconstruction. This suggests that the forensic AUTOPSY Project is best understood not as an isolated classifier, but as one component within a broader computational autopsy ecosystem in which imaging standardization, transfer learning, explainability, auditable workflows, and human-in-the-loop review are recurrent design themes.