Artificial intelligence in forensic pathology: an Australian and New Zealand perspective

The medical school curriculum in Australia and New Zealand offers minimal training in statistics and lacks formal modules on computer science or informatics, leaving graduates with limited foundational knowledge of AI in healthcare. After medical school, junior doctors typically work in hospitals or primary care before specializing.

Forensic pathology, a sub-specialty of pathology under the Royal College of Pathologists Australasia (RCPA), focuses on postmortem examinations. Most practicing forensic pathologists have little expertise in clinical forensic medicine and the forensic pathology curriculum does not mandate education in digital pathology or AI. Consequently, understanding AI in forensic pathology is not a qualification requirement.

Qualified forensic pathologists generally work in government forensic pathology departments or institutes within pathology, health, hospital, or justice sectors. Beyond the RCPA’s annual requirement for continued professional development, teaching, training and research are not mandatory components of their roles (Table 1).

Forensic pathology in Australia and New Zealand operates under a coroner system, with each state and territory governed by its own Coronerʼs Act. Forensic pathologists assist coroners in death investigations but practices vary widely across departments due to differences in legislation and resource availability.

The extent of a postmortem examination, ranging from external examination of the corpse to a full autopsy, is determined by the coroner, informed by advice from the pathologist and other stakeholders. Most departments have access to postmortem computed tomography (CT) scans and digital photography is common, although there are no national standards or guidelines for postmortem imaging. Histology and ancillary tests are typically available, either in-house or outsourced, with test selection guided by the pathologist and existing guidelines.

Paper-based notes and diagrams are still widely used, serving as either formal records or informal references. These are rarely digitalized. Postmortem reports are prepared either through dictation for transcription or typed directly by the pathologist, with formats varying between piecemeal and complete drafts.

Software infrastructure and data storage vary significantly between regions. Departments may use bespoke systems, commercially available software, or borrow systems from parent organizations. Data storage can be regional or cloud-based and not all records are digitalized or centralized.

A formal detailed postmortem report is not always mandatory, especially for cases involving only preliminary examinations. Reporting styles range from free text to template-based formats, often created using word processing software. Once completed, reports are uploaded to the National Coronial Information System (NCIS), which consolidates coronial data to support public health and justice recommendations [56].

The NCIS primarily captures text-based data, including demographics, cause of death and manner of death. It does not store imaging data or detailed postmortem extracts beyond these basics.

When a death is reported, a range of documents are provided which can include a police report, scene information, witness statements, death certificate, special investigations (e.g., sudden infant death report, traffic crash report, diving equipment examination) and the medical history. The level of detail is highly variable and is dependent on the type of death and its surrounding circumstances, the availability at the time of reporting the death and the individual completing the documentation.

Potential applications: adapting from AI applications in the clinical setting, the initial information could be used for case triage, case allocation and turnaround time estimation [62, 66]. For the individual pathologist, this information is used to plan the postmortem examination, such as what type of postmortem examination is needed and what samples and tests are to be undertaken.

Data: the type and quality of data can be variable. Most reports (police, medical or special reports) are commonly a hybrid between template structure and areas of unstructured free text. This information can be historical or contemporary. Witness statements are highly unstructured and akin to a candid script/dialogue and can be in text format or a recording. Scene information is typically unstructured and dependent on the type of death encountered. All of this data collection is performed by humans, human error and bias cannot be avoided. Data collected can be biased (i.e., the amount of detail in a routine case can be vastly different from suspicious ones). Forensic pathologists are only the recipient of this information and data presented and have no control or expertise in deciding the type, method and/or amount of data collected. With local/regional nuances certain data might be captured in one jurisdiction and not in another. Nonetheless, certain key information is commonly manually extracted from these sources and transcribed into a forensic pathology department case management system. What information is considered necessary to extract is department-dependent but commonly includes identifying details (name, date of birth, date and location of death etc.) and case categorization (suspicious, complex, natural etc.).

Current state of research: limited research exists on AI data extraction in forensic pathology. Preliminary studies in medical examiners’ data show promise but require manual editing to address duplication and inaccuracies [6]. There does not appear to be current published research looking at applications of AI data extraction to forensic pathology practice.

Postmortem report generation is variable but frequently involves some element of converting information from photographs, paper autopsy notes and memory to digital text via dictation, self-typing or a transcription program.

Potential applications: assisted diagnostics and workflow.

Commercial AI applications are accessible to forensic pathologists in generating reports. At the most basic level, dictation software packages developed commercially are available (many of which now incorporate AI), and can be either medical (such as Dragon® dictation software, Grundig Business Systems, Nürnberg, Germany) or non-medical-specific. Even with the medical-specific applications, none are tailor-made for forensic pathology and typical forensic dictionaries are not yet available. Many of these applications can be trained to suit the individual doctor as the reporting style of doctors varies with training, background and personal preference [49]. At an advanced level, generative AI in theory could be used to assist in generating comments or suggest relevant pre-existing medical conditions to determine the cause of death from the postmortem findings. In the clinical setting AI has been used to assist in patient notes and hospital admission summaries [38, 68]. Indirectly, when using any computer that interfaces with the internet, pathologists are using AI in some capacity, such as something as simple as using a search engine in researching a disease.

For benchmarking, there are limited data on whether these applications increase efficiency and decrease operational cost in practice, with much of the current research in this rapidly evolving field focused on theoretical advantages [16]. It is also unclear how much AI involvement key stakeholders (coroners, police, lawyers, doctors and family) would accept. For example, dictation, general word processing and basic autopopulated data would probably be acceptable and/or familiar to many people but the use of generative AI to assist formulation of the cause of death and comments based on the information provided for a complex homicide case might not be.

Unlike clinical medicine, forensic pathology significantly interfaces with the law/justice system. It is unclear when a pathologist is asked to provide an opinion as an expert witness to what extent it may be acceptable (even if disclosed) for an AI to assist with data summarizing and the formulation of a second report and medical opinions. A forensic pathologist, or indeed any expert witness, must be able to explain their reasoning in reaching an opinion such that the trier of fact (such as the judge or jury) can assess how this may influence the case outcome [13]. The “black box” of AI does not allow for an understanding of how a given opinion was reached [21] and this remains a shortcoming of explanation in front of a court. Early examples of AI introductions into legal settings (when insufficiently supervised or fact-checked) have had serious consequences secondary to false information generation [42].

Data: Word processor forensic pathology reports and ancillary reports/materials where relevant.

Current state of research: commercial transcription (not specific to forensic pathology) tools are in use but their impact on forensic workflows remains underresearched [30]. Generative AI for diagnostics has seen limited application. Studies, such as AI-based vehicle type identification in pedestrian deaths, show insufficient accuracy for routine forensic use [30].

Once a report has been produced, quality assurance processes such as peer review may be needed.

Potential applications: AI could aid in automated peer review allocation, considering pathologist caseloads, flagged case types (e.g., homicides or pediatric cases) and random inclusion of other cases for quality checks. It can also handle mechanical aspects of peer review, such as detecting data omissions (e.g., measurements, weights), transcription errors, grammar issues and ensuring all referenced materials are included in the digital system. This would streamline the process, allowing pathologists to focus on complex tasks like injury analysis and determining the cause of death. Some tasks are straightforward enough for AI to apply broadly, not just to peer-reviewed cases.

Data: Word processor forensic pathology reports and ancillary reports/materials where relevant.

Current state of research: scant forensic pathology applications. The use of AI has been applied to other medical disciplines for auditing/quality assurance purposes [5, 8, 59], although these systems themselves must be subject to a separate quality assurance process [40]. Research on AI-driven peer review in forensic pathology is sparse.

Postmortem imaging in forensic pathology commonly refers to, but is not limited to, radiology and photography. Most of the AI-based research, albeit limited, in this area is for identification and diagnostic purposes.

Potential applications: assisted diagnostics (including screening for findings that might feed back into case management, such as trauma), injury pattern comparison, postmortem interval estimation and identification.

Existing clinical radiological AI applications in assisted diagnostics may have merit in the postmortem setting, although even current clinical applications remain limited in practice, with much of the existing research focused on a narrow range of applications without external validation [22, 32]. If AI becomes more widely integrated into radiology software, training on postmortem data is likely to be necessary given the significant differences between antemortem CT and PMCT in terms of postmortem changes, body positioning and nature of the pathology (including injuries that are incompatible with life). Digital photography diagnostic applications are similarly limited by the need to train on postmortem data but are also an issue for usefulness of timing as digital photographs are not typically uploaded to a computer until the autopsy is complete. Postautopsy applications as a second-check or safety net could involve functions such as injury classification (such as interpretation and classification of blunt force, sharp force and gunshot injuries and even patterns of formed injuries).

Existing preliminary research in clinical forensic medicine has shown potential of AI to assist in the dating of photographed bruises in living people [60]. Image-recognition applications could enable searching of unlabelled postmortem photographs by organ type or pathology, allowing for content-based image retrieval of relevant images for research, auditing or education purposes that would otherwise require manually searching through thousands of digital files and an even greater number of unrelated images [65].

Data: postmortem CT (PMCT) and postmortem photographs Postmortem magnetic resonance imaging is of limited availability or unavailable to most Australian and New Zealand forensic pathology departments just as in most German language institutions. The use of PMCT is available, or at least accessible, in most Australian and New Zealand forensic pathology departments. The scans are performed by radiographers or trained forensic pathology technicians. Scan protocols may vary between departments. Scanning usually includes at least the head, neck and torso but frequently full body [47]. Clothing may or may not be removed, may be repositioned due to physical constraints and depending on local practice, may seldom be rescanned after clothing is removed and the body is repositioned. Use of PMCT scans is for identification (particularly odontology) and screening purposes (both for safety, such as sharp objects or other hazards and for diagnostics) [46]. On the basis of a pathology sufficiently demonstrated by PMCT (or exclusion of a major pathology, such as in a toxicological death) some departments opt for external only or limited internal examinations [47].

The standard of postmortem photography varies greatly due to training and personal preference. Some pathologists have routine sets of photographs, some only photograph (more or less relevant) pathology and more extensive photographs are typically taken by police photographers if physically present during autopsy [48]. Depending on the department, there may be variability in the camera type, distance, use of flash and use of photographic background.

Current state of research: postmortem research in this field has involved compartmentalized applications in a preliminary research capacity. Odontological applications have been researched for identification (gender and age estimation) purposes, albeit using X‑ray images [1, 45] and there has been research into automated labelling of femoral implants (also potentially useful for identification purposes) [53]. A few applications were investigated in a specific cause of death context including hemopericardium [15] and head injuries [20, 24].

Preliminary studies have also looked at AI-assisted diagnosis of specific pathologies including rib fractures [26] and pericardial effusion [33]. One study showed limited utility of AI in determining the mechanism of skeletal injury [14].

Outside of radiology, digital photograph postmortem imaging AI applications have been less researched but gunshot wound interpretation has been investigated in a preliminary capacity [12, 50]. One preliminary study showed variably accurate application of AI to postmortem interval estimation based on photographed eye opacity [10]. The use of AI showed limited but nonetheless some potential utility in identification of deceased based on postmortem photographs [44]. In relation to the internal autopsy examination, preliminary research has investigated the capacity of AI to identify macroscopic organs from digital photographs [19] but not yet in differentiating (organ-specific) diseases/pathologies.

Histology is used in nearly all internal examinations and varies widely in sampling methods due to differences in training and experience. It helps confirm macroscopic findings, screen for microscopic pathologies and age injuries [35].

Potential applications: assisted diagnostics such as screening for pathology and dating of injuries/pathology.

As originally taken from anatomical/surgical pathology, sampling tissue for histology examination is a common practice in forensic pathology; however, there is variability in the approach to histology, in part due to differences in how postmortem histology is used compared to anatomical pathology. There is little in the way of structured reporting for forensic histopathology and tissues are frequently normal, affected by autolysis, involve organs that are rarely sampled in the living (heart and brain) and provide a cause of death different from autopsy results in only a minority of cases [35, 51]. Cases where a malignancy is directly relevant to death are rare and immunohistochemistry is even less frequently used as well as being affected by postmortem changes.

Potential AI applications in forensic histopathology therefore differ from current AI applications in anatomical pathology that are primarily focused on cancer diagnosis, prognostication and automated tasks, such as mitosis counting [27, 43]. Applications in forensic pathology could include assisted diagnostic dating of injuries and diseases (such as myocardial infarction) as well as screening for specific pathologies, such as histological features of cardiac hypertrophy. Where cancer diagnosis applications are desired, much like in forensic radiology, clinical AI applications are unlikely to be directly transferable (due to factors such as autolysis) and training on postmortem data is necessary.

Data: requires digitalized histology slides. Whilst digital histopathology is only recently becoming more commonplace in Australian anatomical pathology departments [64], the comparative niche field of forensic pathology remains less financially and logistically equipped in remaining up to date with advances in overlapping clinical disciplines. Consequently, even without accounting for in-house AI training on postmortem histology specimens, the costs of digital pathology imaging and storage are magnified compared to the much larger histopathology output of a moderate-sized anatomical pathology department [25, 29]. Alternatives to fully digitalized histology include microscope camera micrographs of regions of interest but these remain limited by the amount of data in terms of magnification, pixelation, depth of field and area of tissue photographed and requires again time and staff to be achieved. Unlike full digital histology, micrographs are an interruption to normal workflow and AI analysis would occur separately instead of in tandem with normal work. Also, with the diminishing number of internal examinations in Australia and New Zealand, the availability of histology slides would consequentially diminish resulting in a lower amount of data for AI training.

Current state of research: whilst only a single study, AI has shown potential in applying an important surgical pathology tool for digital pathology to postmortem tissue in the form of virtual histology staining, including in cases of autolysis [37]. Within the niche of diatom interpretation, studies have shown accuracy in AI-assisted recognition of diatom types [67, 70], although there is no existing research showing that this improves the reliability of the controversial diatom test in diagnosing drowning with certainty. There is very limited research on diagnostic applications of AI in forensic histology, with a preliminary micrograph-based study on myocardial infarction dating [18].

Common ancillary tests in forensic pathology include toxicology, biochemistry, and microbiology (e.g., bacterial cultures and virology). These tests are guided by case history or used broadly in cases with limited information [34].

Potential applications: assisted diagnostics.

These tests (like most diagnostic information) require interpretation in the context of the case history and other findings [13] but nonetheless can provide useful or prompting information in isolation to guide the diagnostics. For example, significantly elevated glucose and ketone levels in cerebrospinal fluid and/or vitreous humor should prompt consideration of a diabetic ketoacidosis, whether or not the history was suggestive of this up to that point. The case context then determines the relevance of such a finding to the cause of death. Whilst more complicated in their interpretation than in the clinical setting due to the effects of postmortem changes, published postmortem reference ranges exist for many of these tests (particularly toxicology and biochemistry) [2, 4]. At their most simple, AI applications in this setting could act as a more sophisticated version of a set of tailored reference ranges, for example, automatically providing a set of comparison non-toxic, toxic and lethal ranges for a given detected drug and matching drugs to their metabolites. More complex functions could include suggested diagnoses, such as the previous diabetes example, grouping central nervous system depressant drugs to suggest a mixed drug toxicity or suggesting that high growth of an identical infectious organism in multiple organs may be more likely to represent a true infection than postmortem overgrowth. These tasks are currently performed regularly by forensic pathologists’ individual knowledge and AI could feasibly decrease the amount of time required to be spent on each task.

Data: simple applications more akin to a traditional computer program than AI may not require much in the way of data if there are established postmortem ranges for comparison. More complicated applications would require training access to both the test results and interpretation of these results in forensic pathologistsʼ reports. As with most potential AI applications in forensic pathology, the lack of standardized reporting is a complicating factor. Furthermore, as these tests are not typically performed on a routine basis, the data are biased towards positive outcomes. Variability in sampling methods (even within the same department), extent of sampling (such as in microbiology), extent of effects of postmortem change and the co-contribution of other findings (such as the clinical history or histology) to the diagnosis mean that sophisticated AI applications in this area are likely to be challenging to train.

Current research: scant research has been done mostly focused on postmortem interval estimation, including a study that showed mixed limited utility of AI in postmortem interval estimation based on microbiota [69] and a postmortem interval study based on vitreous electrolytes which showed improved accuracy over traditional methods [7]; however, challenges, such as environmental effects, body habitus and microbiota variability make AI applications for postmortem tests complex and inherently limited.

The frequent use of the term “potential” in AI discussions underlines the gap between what the technology might achieve and its current limitations. A major issue is AI hallucinations, where generative AI produces false yet plausible information [38, 58]. The use of AI often generates fluent, grammatically correct text, making errors harder to detect [58]. For example, AI has fabricated body mass index (BMI) values without supporting data. In medical documentation, one study found errors in 46% of AI-generated discharge summaries, with 8.4% involving hallucinations [58]. This issue also affects legal cases, where AI-generated fake citations and research papers have been reported. Furthermore, AI models can inherit biases from training data, including racial and gender biases, or lack of representation from certain population groups.

Forensic pathology often involves communication with non-experts, such as coroners, jurors and families. Factually incorrect but well-worded AI output could cause significant harm, especially in legal contexts where research shows that laypeople may trust AI-generated content more than expert documentation [58].

Separate to the issue of AI errors and hallucinations, AI have also been shown to exhibit biases (including racial and gender biases) based on biases in the data they are trained on, or insufficient data from relevant population groups [11, 52].

While AI offers potential benefits for improving workflow and accuracy, its implementation in forensic pathology must be carefully supervised and validated to avoid errors, biases and misuse.

The immediate technical challenges in relation to implementing AI mainly surround the collection, processing and analysis of a large quantity of quality data. This would require initial consultation with computer/data scientists, legal professionals and ethicists, additional hardware and software, change of attitude and increased literacy in AI in the forensic pathology community and standardization of practice with a focus on digitalization of data. After having acquired quality data and provided AI is implemented in the specific areas in question (diagnostics and/or operational), continued benchmarking for its performance is needed in the realm of operational and diagnostic needs and indices.

Biomedical ethics is based on universal principles such as non-maleficence, beneficence, respect for personal autonomy and compliance with legal rules and laws.

The use of AI in forensic pathology would need to consider complex ethical issues. These matters would include but not be limited to impartiality and accountability, avoidance of the uncontrolled use of sensitive data, compliance with privacy and data protection legislation, accessibility of data limited to authorized personnel, application of the principles of informed consent, avoidance of discrimination and human rights violations and transparency of analytical procedures [3, 23, 61].

The use of AI is already present in forensic pathology through tools such as Word processors and search engines. Its integration will grow but forensic pathologists must participate in shaping this transition. Key steps include:

The use of AI offers potential to enhance forensic pathology but success depends on addressing these challenges through strategic planning and ethical monitoring.