Hospitals must manage AI outputs, model versions, access logs and performance over time
US healthcare group issues AI resilience playbook as study identifies commands hidden in image pixels
Series Overview
- September 4 | HIRA Launches First AI-Assisted Claims Review
- September 11 | Medical Imaging AI Platform Competition
- September 18 | Lung Cancer Screening AI and National Follow-Up Care
- September 25 | Breast AI Shows Clinical Value Through Fewer Interval Cancers and Faster Diagnosis
- October 2 | Stroke AI Links Imaging to Treatment and Recovery
- October 9 | DICOM Standardization and Medical Imaging AI Performance and Security
- October 16 | Reimbursement Linked to Clinical Outcomes
- October 23 | Global Market Strategy for Korean Medical Imaging AI
Medical imaging standards connect CT and MRI systems with picture archiving and communication systems and AI analysis. Hospitals must monitor model performance and access records while maintaining clinical operations during security incidents or system failures. / Korea IT Times
Medical imaging artificial intelligence can operate reliably in hospitals only when imaging equipment, picture archiving and communication systems (PACS), radiology information systems (RIS) and electronic medical records (EMR) exchange patient and examination information under consistent rules. Digital Imaging and Communications in Medicine (DICOM), the international standard for medical imaging, provides the foundation for storing and transmitting CT, MRI and X-ray images together with patient, examination and equipment data.
Transmitting images in DICOM format does not automatically ensure the accuracy or security of AI results. If hospitals use inconsistent examination names and anatomical labels, or connect images to the wrong patient record, an AI system may miss a relevant study or send it to the wrong algorithm. Without encryption, access controls, audit logs and model-version management, a standardized network can also spread errors and security threats.
Medical imaging AI performance depends on the algorithm, input data and hospital operating environment. The same product may produce different results as CT manufacturers, reconstruction methods, slice thickness, contrast use, imaging protocols and patient populations change. Hospitals need to know which images and metadata entered an AI system, which model version processed them and who received the results.
DICOM Provides a Common Language for Clinical AI
DICOM is more than a file format. It links patient identity, examination time, body part, imaging equipment and acquisition protocol to the image while defining how healthcare institutions transmit and retrieve the data. DICOMweb extends these functions to web-based environments, allowing systems to search, retrieve and store medical images.
AI-generated information, including the location and size of a lung nodule, a suspected intracranial hemorrhage and a mammography risk score, must remain accurately linked to the source image and patient examination. If results appear only in a separate application or static image file, radiologists may struggle to compare them with the original study or track the same lesion during follow-up. Structured DICOM objects allow measurements, annotations and source images to be managed within the existing reading environment.
Using the same standard does not guarantee interoperability between products. Vendors may support different DICOM functions, data fields and methods for displaying AI results. Before deployment, hospitals should review the vendor’s DICOM conformance statement and test image transmission between the product, imaging equipment and PACS.
Inconsistent examination names are another source of error. The same chest CT may be recorded under different names and codes across hospitals and scanners, making it difficult for an AI system to identify the correct study. Systems that depend on basic text-matching rules may miss an emergency examination or analyze an image outside the product’s intended use. Hospitals therefore need standardized information on the body part, imaging method, contrast use and image series, along with procedures for handling exceptions.
Hospitals must also verify how AI results appear on clinicians’ worklists. If a stroke or intracranial hemorrhage system moves a suspected case to the top of the queue but fails to deliver the alert to the responsible physician, treatment will not begin sooner. Repeated false-positive alerts can also cause alert fatigue and delay the review of other emergency cases.
Stable integration with PACS, electronic medical records and multi-institutional workflows could become a more important competitive factor than the accuracy of a single algorithm,” said Youna Kim, deputy general manager of CorelineSoft’s media team. In markets such as Germany’s lung cancer screening program, where first and second readings, prior-image comparison, follow-up and quality management are connected, companies must support the full screening process rather than a single examination.
CorelineSoft said demand is increasing for AI systems that connect multiple institutions and readers as Germany introduces low-dose CT lung cancer screening under its statutory health insurance system. However, contracted sites, completed installations, examination volumes and clinical adoption rates measure different stages of deployment. National screening programs should therefore assess AI analysis volumes, prior-image matching, interpretation times, review rates and adherence to follow-up examinations.
The effect of a system failure also increases as AI becomes more deeply embedded in hospital operations. Images must not disappear from reading worklists when a cloud connection fails or a vendor’s server becomes unavailable. Clinicians should be able to continue reading examinations without AI and identify missed studies and affected patients after service is restored.
Security Must Extend from Image Transmission to Clinical Recovery
The DICOM standard includes security mechanisms supporting encrypted communication and audit records. Transport Layer Security can protect data in transit, while audit trails can record who accessed a patient’s information, when the access occurred and what action was taken. These capabilities provide little protection if a hospital does not activate them or if legacy imaging equipment does not support the relevant security profiles.
Healthcare institutions should verify encryption, user authentication, access privileges, certificate and cryptographic-key management, and log-retention periods for each product. Contracts should define when AI vendors and cloud operators may access patient images. Remote-maintenance accounts and external network connections should remain open only for the required period and purpose, with every access recorded.
On October 8, the US Health Sector Coordinating Council Cybersecurity Working Group released its Playbook for AI Clinical Resilience. It calls for clinicians, IT and security teams, governance leaders and compliance officers to coordinate their response from the detection of an AI-related threat or failure through patient-safety review, regulatory reporting and system recovery.
The playbook treats medical imaging AI security as more than preventing data leaks. Healthcare institutions should decide in advance how clinicians will return to established workflows when AI assigns an incorrect priority or an emergency alert stops working. They also need procedures for identifying affected patients, reviewing potentially incorrect results and deciding when the system can safely resume operation. Although the playbook is not legally binding, it treats AI failure as a clinical-continuity and patient-safety issue.
Researchers have also identified a new attack method. In a study published by Scientific Reports on October 3, investigators embedded malicious instructions in the pixels of dental panoramic radiographs and tested four vision-language models: GPT-4o, Gemini 2.5 Flash, Claude Sonnet 4.5 and MedGemma 4B.
The researchers conducted 58,320 inference calls using 270 images and found that all four models were affected. Under some conditions, the attack success rate reached 62.6%. An optical character recognition-based defense that detected and removed text embedded in the image reduced the overall attack success rate from 15.6% to 0.2%.
The findings show that protecting image headers and transmission networks may not prevent commands embedded in pixels from manipulating a multimodal AI system. Hospitals using vision-language models to process images received from outside institutions or uploaded by patients should verify both the image source and any text contained within it. If the system detects an abnormal instruction, it should stop the AI analysis and refer the case for human review.
AI Performance Requires Continued Oversight After Standardization
The study does not establish that every medical imaging AI product currently used in hospitals faces the same level of risk. It was a controlled adversarial experiment using a public dental-image dataset, not an investigation of a breach at a healthcare institution. Conventional imaging AI designed primarily to detect lesions has a different attack surface from a vision-language model that interprets written instructions.
The UK government accepted all 44 recommendations of the National Commission into the Regulation of AI in Healthcare on October 6. The government plans to manage medical AI through a framework covering deployment, real-world performance, clinical responsibility, patient transparency and the full product lifecycle. It concluded that existing medical-device regulations do not adequately reflect the speed of AI development and said premarket review and postmarket monitoring should be connected.
Nabile Safdar, MD, chair of the Society for Imaging Informatics in Medicine, has described the responsible use of imaging AI as “an ongoing process rather than a single event.” He said dedicated teams must consistently apply defined procedures supported by appropriate methods and technology, identifying interoperability standards, expert participation, AI science and methodological rigor as foundations for clinical implementation.
Safdar’s assessment shows why oversight must continue after DICOM integration. Hospitals still need to manage model versions, performance changes, access logs, errors and criteria for suspending use. Standardization provides the connection to hospital systems, while institutional monitoring and governance determine whether the technology remains safe.
South Korea can use its extensive PACS infrastructure, national health insurance system and screening data to build a national performance-management framework for medical imaging AI. This would require links between product approvals and change-management information held by the Ministry of Food and Drug Safety, hospital usage records, equipment and imaging conditions, and clinical data held by the Health Insurance Review and Assessment Service. The scope of data collection and access rights must also protect patient information and commercially sensitive material.
AI procurement contracts used by hospitals and public institutions should specify more than whether a product supports DICOM. They should identify the supported data objects and examination codes, formats used to store AI results, encryption and logging requirements, model-update procedures, response times for outages and thresholds for performance deterioration. They should also define how data will be handled and how another product can replace the system when the service ends.
Standardization does not require every medical imaging AI product to operate in the same way. It creates a consistent structure linking the patient, examination, image and AI result so that hospitals can identify the cause of an error and maintain clinical records and follow-up care when a product is replaced. Medical imaging AI can function as core hospital infrastructure only when standardized connectivity, performance monitoring, cybersecurity and recovery procedures operate together.
Part 7 will compare medical imaging AI reimbursement in the United States and South Korea and examine the clinical and economic evidence required to move from regulatory approval and billing codes to actual insurance coverage.
