Artificial intelligence-assisted chest radiograph interpretation in Role 2 military field hospital settings: a controlled experimental study

Abstract

Introduction

Forward military field hospitals often operate in battle zone environments where access to specialized personnel, such as radiologists, is limited, complicating the accuracy of diagnostic imaging. Chest radiographs are crucial for assessing thoracic injuries and other conditions, but their interpretation frequently falls to non-radiologist personnel. This study evaluates the effectiveness of an artificial intelligence (AI)-assisted model in enhancing the diagnostic accuracy of chest radiographs in such resource-limited settings.

Methods

Nine board-certified military physicians from various non-radiology specialties interpreted 159 anonymized chest radiographs, both with and without the support of AI. The AI model, INSIGHT CXR, generated automated descriptions for 80 radiographs, whereas 79 were interpreted without AI support. A linear mixed-effects model was used to assess the difference in diagnostic accuracy between the two conditions. Secondary analyses examined the effects of radiograph type and physician specialty on diagnostic performance.

Results

AI support increased mean diagnostic accuracy by 9.4% (p<0.001) from pretest to post-test, representing a 23.15% relative improvement. This improvement was consistent across both normal and abnormal findings, with no significant differences observed based on radiograph type or physician specialty. These findings suggest that AI tools can serve as effective support in field hospitals, improving diagnostic precision and decision-making in the absence of radiologists.

Conclusions

This study highlights the potential for AI-assisted radiograph interpretation to enhance diagnostic accuracy in military field hospitals. If AI tools are proven reliable, they could be integrated into the workflow of forward field hospitals, improving the quality of care for injured personnel. Immediate benefits may include faster diagnoses, increased personnel readiness, optimized performance, and cost savings, leading to better outcomes in combat operations.

Level of evidence

II. Diagnostic Test.

WHAT IS ALREADY KNOWN ON THIS TOPIC

• In military field hospitals, chest radiographs are essential for assessing thoracic injuries. However, due to the limited availability of radiologists in these settings, non-radiologist personnel often interpret these images, potentially impacting diagnostic accuracy.

WHAT THIS STUDY ADDS

• This study demonstrates that artificial intelligence (AI)-assisted interpretation significantly improves diagnostic accuracy for chest radiographs in a field hospital environment, achieving a 9.4% increase in accuracy regardless of the radiograph type or physician specialty.

HOW THIS STUDY MIGHT AFFECT RESEARCH, PRACTICE OR POLICY

• The integration of AI tools in military field hospitals could enhance diagnostic reliability and speed in resource-constrained environments, leading to quicker and more accurate clinical decisions and potentially establishing AI as a standard diagnostic aid in similar settings.

Background

The delivery of healthcare in field hospital (military treatment facility) conditions presents unique challenges, particularly due to the often limited availability of specialized medical personnel and diagnostic resources. In the context of NATO (North Atlantic Treaty Organization) military operations, facilities such as the Role 2 Forward (Role 2F, formerly Role 2 Light Maneuver) and Role 2 Basic (Role 2B) field hospitals are designed to provide immediate surgical care to combat casualties.^{1 2} Chest radiography is a fundamental diagnostic tool used in emergency settings within Role 2 facilities, providing rapid imaging essential for the assessment of thoracic injuries and other critical conditions in combat casualties. However, Role 2 field hospitals typically consist of a surgical team and do not include radiologists among their assigned medical personnel.² Consequently, the interpretation of chest radiographs falls to non-radiologist medical staff, such as surgeons and anesthesiologists. This reliance on non-specialist interpretation necessitates that these medical officers apply their general diagnostic skills to radiographic analysis, which may present challenges in accuracy and efficiency under the high-pressure, resource-limited conditions of a field hospital. This practice can result in diagnostic inaccuracies, either through the underestimation or overestimation of the severity of radiographic findings, thereby potentially compromising patient care.

The advent of clinical applications of machine learning (ML) in medical diagnostics offers a promising solution to this problem.³ These artificial intelligence (AI) algorithms are most often based on deep learning techniques, specifically convolutional neural networks (CNNs), and more recently, vision transformers employing attention mechanisms.⁴ They have demonstrated considerable proficiency in image classification, object detection, and the interpretation of medical images, including those that are blurred or otherwise challenging to analyze.^{5 6} These capabilities render ML-based tools particularly suitable for the automatic description of plain radiographs in environments where radiological expertise is unavailable.⁷

This study aims to assess the usability and effectiveness of an automatic chest radiograph description model in simulated field hospital conditions. Specifically, the project investigates whether an ML model can enhance the diagnostic accuracy of trained medical professionals, in scenarios where a radiologist is not available. By evaluating the performance of the model alongside medical practitioners, the study seeks to analyze the viability of integrating such technology into the operational workflow of field hospitals. In the discussion section, we also briefly explore the cybersecurity and legal implications associated with implementing this technology.

The successful implementation of this technology could significantly enhance the quality of medical care in combat operations by providing reliable diagnostic AI support and mitigating the limitations posed by the absence of specialized radiologists. Consequently, the findings from this study could pave the way for the broader adoption of automated diagnostic tools in military and other resource-limited healthcare settings.

To the best of our knowledge, this is the first study to rigorously evaluate the role of AI support in radiological decision-making within the context of a military field hospital.

Methods

Participants

This study was conducted with the participation of nine physicians from various clinical specialties, including anesthesiologist–intensivists, general surgeons, and orthopedic–trauma surgeons (see table 1). All participating physicians were board-certified specialists and active-duty military personnel (professional experience: 8–19 years; military deployment: 0–18 months).

Table 1. Characteristics of participating physicians.

ID	Clinical specialty	Clinical experience (years)	Deployment experience (months)	Active duty
1	Anesthesiology–intensive care	19	18	Yes
2	Anesthesiology–intensive care	12	0	Yes
3	Anesthesiology–intensive care	8	6	Yes
4	Anesthesiology–intensive care	11	11	Yes
5	Anesthesiology–intensive care	16	0	Yes
6	Orthopedics–traumatology	11	4	Yes
7	Orthopedics–traumatology	10	0	Yes
8	General surgery	8	0	Yes
9	General surgery	9	0	Yes

Dataset and image handling

A total of 159 chest radiographs were selected by a radiologist (vs) from patients examined at the Military University Hospital Prague. To establish the ground truth, these images were independently interpreted by two board-certified radiologists. In cases where the two radiologists had discordant interpretations, a consensus was reached through joint discussion to establish the ground truth. The pathological findings of interest included atelectasis, consolidation, pleural effusion, lung lesions, cardiomegaly, and pneumothorax. Of the 159 selected radiographs, 70 showed normal findings, whereas the remaining 89 demonstrated varying numbers of pathological findings. Specifically, 63 radiographs had a single pathological finding, 21 showed two distinct pathological abnormalities, and 5 displayed three separate pathological findings. All images, stored in Digital Imaging and Communications in Medicine (DICOM) format, were anonymized prior to analysis by the third party to ensure patient confidentiality.

The chest radiographs were divided into two groups: the first group had images accompanied by AI descriptions generated by Lunit INSIGHT CXR (Lunit, Seoul, South Korea), whereas the other group had images without the AI descriptions. The assignment of radiographs into these two groups was conducted using stratified random sampling to ensure a balanced distribution of normal and abnormal findings in both groups (p=0.570). The order in which the radiographs were presented to the participants was randomized, but the sequence remained consistent across all participants to control for order effect and practice effect (see figure 1).

Figure 1. Schematic overview of the experimental design and methodology. AI, artificial intelligence.

Machine learning model

The software used in this study was INSIGHT CXR, developed by Lunit (Seoul, South Korea). INSIGHT CXR is a neural network-based software intended to assist radiologists in interpreting anteroposterior chest radiographs by detecting and localizing common radiologic findings.

The INSIGHT CXR system has received Food and Drug Administration 510(k) clearance for use in clinical settings, ensuring it meets regulatory standards for safety and efficacy, allowing the software to be used in healthcare institutions across the USA.

The model operates using deep learning algorithms, specifically CNNs, trained on a large dataset of over 3.5 million manually annotated chest radiograph images. The model is designed to detect ten prevalent abnormalities found in chest radiographs (lung nodules, pneumothorax, pleural effusion, consolidation, atelectasis, pneumoperitoneum, cardiomegaly, mediastinal widening, calcification, and fibrosis). The CNN consists of multiple overlapping feature detectors organized hierarchically in layers, where lower layers focus on detecting basic features like edges and textures, and higher layers identify more complex patterns, such as shapes of specific pathological findings. The input radiograph is analyzed at the pixel level, generating a probability score for each abnormality. This results in a heatmap, which can be converted into a lesion contour based on a defined probability threshold. The probability score ranges from 0 to 100, with the developer considering a threshold of 15 as clinically significant (See figure 2).

Figure 2. AI-assisted chest radiograph interpretation using Lunit INSIGHT CXR, featuring a heatmap overlay highlighting areas of abnormality. The intensity of the heatmap indicates the likelihood of pathological finding. The accompanying report lists an overall abnormality score and specific abnormality score for each pathological finding. AI, artificial intelligence.

Participant evaluation process

Participants were tasked with interpreting the radiographs based on their medical expertise. No specific training was provided in chest radiograph interpretation or the use of Lunit INSIGHT CXR prior to the testing. A web-based interface was developed specifically for this study. Participants sequentially interpreted radiographs in their original quality using a DICOM viewer embedded within the interface. In cases where AI support was provided, a second image was presented, displaying the original radiograph overlaid with a heat map highlighting areas of potential abnormalities identified by the AI. Participants could toggle freely between the original image and AI output before submitting their final interpretation. Additionally, numerical probability scores for specific findings were provided to assist participants in their interpretation (see figure 1). Participating physicians did not receive any specific training in the use of AI assistance for radiograph interpretation.

The participants were blinded to any patient metadata to avoid bias. Each physician assessed a total of 159 radiographs (80 radiographs with AI descriptions and 79 without AI descriptions, randomly distributed). This allowed for a direct comparison of their diagnostic accuracy with and without the support of the AI.

Outcome measures

Primary outcome:

• The difference in diagnostic accuracy (proportion of correctly interpreted radiographs) with and without AI support, measured against the ground truth.

The primary metric for evaluating the effectiveness of the AI (ML model) was the diagnostic accuracy of physicians, both with and without AI support. This was measured by comparing the physicians’ interpretations of radiographs with the ground truth. Each interpretation was classified as correct (score 1) if, and only if, all the true outcomes were identified, and as incorrect (score 0) otherwise. We also conducted a sensitivity analysis using an alternative scoring method, categorizing interpretations as: (1) correct (1—full match between identified and true outcomes); (2) partially correct (0.5—at least some of the true outcomes identified); and (3) incorrect (0—none of the true outcomes identified). Since the significance tests for both scoring methods resulted in identical decisions regarding hypothesis rejection, we present the results using the dichotomous scoring (correct vs incorrect) for simplicity. This approach also allows for a straightforward interpretation of the mean score as the percentage of correctly identified radiographs.

Secondary outcomes:

• Diagnostic accuracy stratified by type of radiograph findings (normal vs other).

• Diagnostic accuracy stratified by clinical specialty of the physicians (anesthesiologist-intensivist vs others).

• Sensitivity and specificity of the Lunit INSIGHT CXR software at various probability score thresholds.

Statistical analysis

Basic descriptive statistics, including measures of central tendency and variability, were used to characterize the sample participants and the success of chest radiograph interpretation with and without AI support. A linear mixed-effects model was applied to study differences in the radiograph interpretation success with and without AI support. In these models, repeated assessments of different radiographs under experimental (with AI support) and control (without AI support) conditions were nested within participants, and the compound symmetry type of covariance matrix was used to model the relationships among the multiple measurements. We next extended the model to examine the effects of AI support for type of radiograph finding (normal vs other) and clinical specialty of participants (anesthesiologist–intensivist vs others). The diagnostic performance of the Lunit CXR software was evaluated using receiver operating characteristic (ROC) analysis. In the ROC analysis we calculated the area under the curve (AUC), sensitivity and specificity at the following probability thresholds: (1) 15 (a default setting by the software provider), (2) 30, (3) 50 and (4) 75. Statistical analyses were performed using IBM SPSS V.25.0 (Chicago, Illinois, USA).

Results

Interpretation of radiographs with AI support was more successful among all participants, with improvement ranging from a low of 0.7% to a high of 17.2%. A linear mixed model revealed that the AI support led to a statistically significant increase in correct interpretation of radiographs from 40.6% to 50.0% (p<0.001; see figure 3).

Figure 3. The comparison of the probability of correct evaluation of chest radiographs between two conditions: without AI support (without AI) and with AI support (with AI). The y-axis represents the probability of correct evaluation, and lines connect individual physician performances across both conditions. Error bars indicate variability in the results. AI, artificial intelligence.

AI support was similarly effective regardless of the type of findings (normal vs other; interaction effect p = 0.199) and the clinical specialty of the participants (anesthesiologist–intensivist vs others; interaction effect p = 0.369), as indicated by non-significant interaction effects.

Radiograph interpretation success was however significantly (p<0.001) higher for normal findings compared with other findings with a mean difference of more than 52%. The anesthesiologist–intensivists were slightly more successful (7.6%, p=0.040) when compared with other clinical specializations.

The ROC analysis revealed that the AUC for the Lunit INSIGHT CXR software using the default probability threshold (ie, 15) was non-significant (AUC=0.500, p=1.000, sensitivity=100%, specificity=0%). The overall diagnostic performance of the Lunit CXR increased with an increasing probability threshold and the highest AUC was observed at the highest tested threshold value of 75 (AUC=0.727, p<0.001, sensitivity=84.6%, specificity=59%)—see figure 4.

Figure 4. ROC curve for AI-assisted diagnostic performance illustrates the sensitivity versus 1-specificity for different Lunit CXR score cut-off thresholds (default=15, 30, 50, and 75). The performance of each threshold is compared against the reference line, with diagonal segments produced by ties. The y-axis represents sensitivity, and the x-axis represents 1-specificity. AI, artificial intelligence; ICT, Information and Communication Technology; FDRS, field-deployable radiology system; ROC, receiver operating characteristic.

Discussion

Principal findings

The present study demonstrated that AI support significantly improved the correct interpretation of chest radiographs, with mean diagnostic accuracy increasing by 9.4% (p<0.001), representing a 23.15% relative improvement. This finding is consistent with the growing body of literature suggesting that AI can enhance diagnostic processes across various medical imaging modalities.⁸,¹² Participating physicians did not receive prior training in the use of AI assistance for radiograph interpretation, which suggests that with adequate training, the improvements in diagnostic accuracy could have been even more pronounced. The relatively modest overall diagnostic accuracy may, in part, be due to the strict definition of correctness—requiring complete agreement with the ground truth across all 10 evaluated diagnostic parameters—which may not fully reflect the algorithm’s potential utility in real-world clinical settings.

One of the strengths of our study design is that, unlike many previous studies that directly compare physician performance to AI systems,¹³,¹⁷ our research focuses on how AI support enhances diagnostic accuracy when used in collaboration with physicians, emphasizing the synergistic potential of human-AI collaboration.

The consistency of AI’s effectiveness observed across different radiograph findings and physician specialties suggests a broad applicability of AI tools in diverse clinical scenarios. Notably, the higher success rates for normal findings align with previous reports indicating that AI algorithms are particularly adept at recognizing patterns consistent with the absence of pathological finding.^{13 18 19} This capability could prove invaluable in high-throughput settings where the rapid triage of normal studies could expedite patient care.

The slight advantage of anesthesiologist–intensivists in diagnostic accuracy may be attributed to their extensive experience with critically ill patients and familiarity with chest imaging. However, the relatively modest difference suggests that AI assistance could help level the playing field. Pham et al found that AI support reduced interobserver variability among physicians interpreting imaging studies, promoting more consistent diagnostic outcomes.²⁰ This indicates that AI tools can enhance diagnostic confidence and accuracy across various specialties.

The default probability threshold for Lunit INSIGHT CXR is set by the developer at 15 out of 100. Our ROC analysis showed that this lower threshold prioritizes sensitivity, making it less likely for significant pathological findings to be missed. This setup allows the AI to filter out normal findings with high confidence, helping prioritize images that likely contain abnormalities. Yoo et al demonstrated that further lowering this threshold below 15 for triage purposes removed 41.6% of normal images from the worklist without missing visible abnormalities using the same software.²⁰ Conversely, in our study, progressively increasing the threshold improved specificity, with the highest AUC (0.727) observed at a threshold of 75.

The role of AI in military medicine

Incorporating AI into the modern military is not just a technological advancement but a strategic imperative. The US national strategy heavily emphasizes the integration of AI into warfare, recognizing its potential to transform military operations and enhance decision-making processes. The 2020 National Defense Authorization Act underscores this priority, mentioning AI explicitly multiple times, reflecting its crucial role in future combat readiness. Moreover, the National Defense Strategy highlights the importance of leveraging commercial breakthroughs in AI and ML to maintain a competitive edge, ensuring that the military can operate with increased precision, efficiency, and adaptability in complex environments.²¹

In a military field hospital, where rapid and accurate diagnosis is critical under resource-constrained conditions, the integration of AI for imaging interpretation presents transformative potential. Although this example focuses on radiograph interpretation, AI can also be applied to other imaging modalities and laboratory results, enhancing diagnostic capabilities across a range of tests and analyses in environments with limited resources. AI can rapidly and precisely analyze medical images, flagging abnormalities such as pneumothorax or pleural effusion, which allows clinicians to prioritize urgent cases without waiting for external consultation. This capability is particularly beneficial in combat scenarios where time-sensitive decisions are crucial for patient outcomes.

The implementation of AI in this environment also aligns with the operational needs of forward medical roles, which must remain mobile and responsive. Once integrated into the clinical workflow, AI systems can operate autonomously, reducing reliance on external resources and allowing for continuous operation even in isolated or communication-compromised settings. Moreover, AI can help alleviate the cognitive load on clinicians, who often must make rapid decisions under pressure, by providing a preliminary analysis that can guide their interpretations and actions.

If AI-enhanced diagnostic technologies prove effective, their implementation could be expanded to all levels of field medicine, including pre-Role 1 settings, where AI could assist combat medics in diagnosing conditions such as bleeding and pneumothorax using ultrasound or potentially emerging technologies like impedance tomography or microwave imaging. In Role 1 and Role 2 settings, AI could support physicians and medics facing similar diagnostic challenges. In Role 3 and Role 4 medical facilities, where general radiologists are present but subspecialists (eg, neuroradiologists) may not be available, AI could play a complementary role in helping bridge the expertise gap. AI tools specifically designed to interpret complex imaging, such as brain CT or MRI scans, can, in the future, assist general radiologists by providing preliminary analyses and highlighting areas of concern, allowing them to focus more on interventional procedures.

However, the integration of AI also presents challenges, such as ensuring interoperability with existing military medical systems and training personnel to effectively use these tools. Continuous validation and monitoring of AI performance in real-world conditions are necessary to maintain its reliability and safety. Despite these challenges, AI’s ability to enhance workflow efficiency, particularly in terms of reducing diagnostic delays and improving the accuracy of image interpretation, suggests that its implementation could significantly enhance the clinical capabilities of military field hospitals like Role 2.

Telemedicine

Although AI support is a new and rapidly advancing technology, telemedicine has long served as a proven method for providing remote expertise in military field hospitals. Telemedicine in radiology leverages communication technologies to allow remote radiologists to interpret medical images and provide expert guidance to clinicians in military field hospitals. This approach ensures access to specialized expertise in environments where radiologists are not physically present.²² Compared with AI, telemedicine offers the advantage of human judgment and adaptability, facilitating nuanced decision-making in complex clinical scenarios. However, its reliance on stable communication links and security risks (see below) can be a significant drawback in combat zones. AI, on the other hand, provides rapid, automated analysis without the need for real-time communication.

Although AI and telemedicine both offer valuable solutions in military medical imaging, directly comparing them is challenging due to their distinct approaches, and the literature on this topic remains limited. The study by Huang et al explored the use of AI for interpreting chest radiographs in an emergency department setting. The study compared AI-generated reports with those produced by in-house radiologists and teleradiology services across 500 patient cases. The findings demonstrated that the AI-generated reports were of similar clinical accuracy to those of radiologists and superior in textual quality compared with teleradiology reports. This suggests that AI could serve as a valuable tool in settings with limited radiology resources, providing timely and accurate imaging interpretations that could aid in clinical decision-making.²³

Cybersecurity considerations

Data transfer in military field environments is critical for telemedicine applications and ensuring continuity of care when patients are transferred via strategic evacuation (STRATEVAC) to higher-role hospitals or domestic facilities. The transfer of radiological data between Role 2, Role 3, and Role 4 facilities in combat environments is associated with not only technological challenges but also significant security risks. Radiological equipment used for diagnosis in field conditions may emit detectable electromagnetic waves or other radiation, which adversaries could exploit for navigation or targeting.²⁴ Based on experiences from the conflict in Ukraine, where Russian forces have employed advanced surveillance technology and drones, it is evident that radiological equipment on the battlefield or at Role 2 must be carefully camouflaged and protected from detection. Recommended countermeasures include technologies and tactics such as camouflage, deception, use of underground structures, and enhanced air defenses,²⁵ including the placement of emergency hospital facilities outside the direct range of adversary detection systems.²⁶ In addition to physical risks, cyber attacks targeting data transmission between roles are a growing concern.

From a user perspective, integrating an AI application directly into a device connected to an X-ray diagnostic tool (field-deployable radiology system, FDRS) on the battlefield for on-site data analysis seems ideal. However, this setup poses cybersecurity challenges, especially in creating a device that combines medical functionality with computing capabilities, increasing the likelihood of vulnerabilities. To mitigate this, it is advisable to separate the devices and limit data transfer to essential communications between them.

Maintaining high levels of cybersecurity and resilience for the equipment used in creating radiological images, and particularly for those facilitating their processing and transmission to Role 4, is essential. The following basic principles should be applied to ensure cybersecurity:

1. Isolation of FDRSs and applications via air gaps.

2. Updates sourced exclusively from military-verified platforms (preferably offline).

3. Encryption of the FDRS computer system and data storage, with strong encryption standards such as AES (Advanced Encryption Standard, 256-bit key) for data “at rest”.

4. Use of encryption for data transfer between devices, employing protocols like Secure Sockets Layer/Transport Layer Security for secure communication.

5. Use of virtual private networks for remote access or data transfers between sites (Role 2 to Role 4) to ensure encrypted traffic.

Thus, we propose a system for the transmission and storage of radiological images in low-connectivity environments that combines offline and semioffline solutions, enabling efficient data collection and distribution across military medical service levels (see figure 5). This approach ensures temporary storage of radiology images until a secure connection to a central network is available, at which point the data can be synchronized and analyzed by augmented AI systems or through telemedicine with a radiologist.

Figure 5. The diagram illustrates the process of generating, evaluating, and securely transmitting radiological data from the Role 2 Forward (Role 2F) to higher military medical echelons, specifically Role 3 and Role 4. Within the Role 2F is a field-deployable radiology system (FDRS), which comprises a mobile X-ray machine, specialized imaging software, and a computer system equipped with artificial intelligence (AI) for image analysis, offline data backup, and data encryption. This system allows for on-site preliminary diagnostic evaluations, providing immediate INSIGHTs even in the absence of a radiologist. The FDRS enables offline analysis through autonomous AI, ensuring that diagnostic capabilities are maintained even in disconnected environments. To safeguard sensitive medical and operational data, encryption can be applied locally at the computer system level within the FDRS. This ensures that data remains protected, even in the event of physical interception by adversaries. However, whereas local encryption offers robust security, it presents challenges related to the distribution of encryption keys to Role 2 and beyond. A possible compromise is combining local encryption of the operating system (OS) with secure data transmission methods, such as asymmetric encryption (eg, PGP, S/MIME), ensuring that data remains encrypted in transit without overburdening the distribution of encryption keys. This hybrid approach enhances both local and remote data security, ensuring that sensitive information remains protected throughout the transmission process, even when the FDRS system is compromised. AUC, area under the curve; PGP, Pretty Good Privacy; S/MIME, Secure/Multipurpose Internet Mail Extensions.

The proposed solution emphasizes cybersecurity by leveraging encryption and controlled data access methods to minimize potential threats. Secure offline and online storage and transmission of data reduce the risk of interception, enhancing the protection of both military personnel and sensitive patient data. Furthermore, encryption of the device’s operating system is critical in case the equipment falls into adversary hands.

Legal considerations

Several legal issues arise from the use of diagnostic AI in clinical practice. The specific answers to relevant questions may vary depending on each legal system; however, their fundamental principles are generally similar.

Every medical intervention must be performed in accordance with the standard of care. The use of diagnostic AI is gradually becoming an acknowledged part of medical procedures in several medical fields such as cardiology or radiology. In general, it would be wrong to consider medical AI a deviation from professional standards; it is well possible that in the foreseeable future, these standards will mandate the use of AI in many procedures. However, the efficacy and safety of each new application need to be proven before it can be considered part of the standard of care.

The standard of care inherently involves consideration of the circumstances of a particular case (nobody is obliged to do what is impossible²⁷). Therefore, the use of AI in a combat zone may represent a safer option for providers than its use in standard circumstances. In cases where only non-radiologist specialists (or other healthcare professionals) are available, diagnostic AI may represent the safest and most efficient solution, even if it has not yet matched the outcomes of board-certified radiologists. Even if compliance with the standard of care in a specific case is questionable, using AI as a last resort to save a patient’s life or health may be legally defensible, as it would be performed in the state of necessity.

The majority of AI solutions used in healthcare are delivered under contracts where the supplier acts as the licensor of the software. It is important to note that if a third party, typically the software supplier, occasionally has access to data collected by the software user (eg, for software updates or remote maintenance), this third party becomes a data processor. In such cases, the healthcare provider is the data controller, whereas the supplier of the AI solution acts as the data processor. The controller is responsible for implementing organizational and technical data security measures. Therefore, they must have a comprehensive overview of the life cycle of the personal data being processed. It is essential to map the process of generating personal data—such as images and descriptions—and understand how the data flows through various infrastructure components. This includes identifying where the data will be stored, the repository where it will be processed, how long it will be archived, and the infrastructure path it will follow. A clear understanding of these data flows is necessary for the data controller to establish appropriate obligations for the processor and other involved entities within the contractual framework, which must be stringent due to the sensitivity of the data. Additionally, measures must be taken to prevent cybersecurity incidents, as these can compromise the integrity of sensitive data and pose significant security risks. Addressing issues such as proper access controls, permissions policies, and, where necessary, segregating parts of the infrastructure into a sandbox environment or entirely separate servers can help mitigate these risks.

Limitations

This study has several limitations that should be acknowledged. First, the number of participants was limited to nine board-certified specialists, which may restrict the generalizability of our findings. A larger sample size could provide more robust data and potentially uncover variations in performance across different levels of experience and expertise. Future studies should aim to include a broader and more diverse group of participants to better evaluate the model’s effectiveness across different clinical settings.

Second, the ML model used in this study was capable of detecting only a limited number of pathological findings. Expanding the model’s capability to detect a broader range of pathological findings would be essential for its application in more comprehensive clinical environments.

Future directions

Future improvements should focus on expanding the model’s diagnostic capabilities to include a wider range of pathological findings, leveraging more diverse datasets and advanced ML techniques. New models are emerging with the capability to interpret more complex forms of medical imaging, such as brain CT scans. Enhancing integration into clinical workflows is also crucial; developing intuitive interfaces and embedding AI tools within electronic health record systems could streamline diagnostics and improve efficiency. Additionally, targeted training programs are needed to equip clinicians with the skills to effectively interpret and use AI outputs. Finally, longitudinal studies should assess the long-term impact of AI on clinical practice, ensuring that these tools contribute meaningfully to patient care.

Conclusion

The results of this study demonstrated a statistically significant 9.4% improvement in diagnostic accuracy with AI assistance, suggesting potential value in supporting medical decision-making during combat operations. Although these findings are encouraging, they should be interpreted in the context of the study’s limitations and the modest overall performance observed.

If automatic interpretation of medical imaging continues to improve in accuracy and reliability, such technology could play a supportive role in forward field hospitals by enhancing diagnostic confidence under resource-constrained conditions.

In addition to potentially improving the quality of care, AI integration in military medical settings may offer several operational advantages. These include reducing operational costs, improving personnel readiness, and supporting clinical performance. By easing the diagnostic burden on non-specialist medical personnel, AI tools could assist in streamlining decision-making processes and contribute to more efficient healthcare delivery in combat and other austere environments.

Data availability statement

Data are available upon reasonable request.