Medical Imaging AI

Executive Summary

Medical imaging AI — the application of computer vision and deep learning to radiology, pathology, and other image-based diagnostic modalities — represents the most mature category of clinical AI, with hundreds of FDA-cleared products and a decade of published clinical evidence. For architects, medical imaging AI is distinctive in three ways: it operates on DICOM images rather than text, it almost always requires FDA regulatory pathway engagement as a Software as a Medical Device (SaMD), and its integration with the clinical workflow occurs through Picture Archiving and Communication Systems (PACS) and radiology workflows rather than EHR-direct integrations. This chapter covers DICOM fundamentals, radiology AI integration patterns, the FDA 510(k) pathway for imaging AI, and the enterprise architecture for deploying FDA-cleared imaging AI alongside an existing radiology workflow.

Learning Objectives

After reading this chapter, you will be able to:

  • Explain the DICOM standard and identify the DICOM objects relevant to radiology AI integration
  • Design the integration architecture for a radiology AI product within a PACS-centered radiology workflow
  • Identify the FDA regulatory pathway requirements for radiology AI products
  • Evaluate a radiology AI product's regulatory clearance against its intended use at a healthcare organization

Business Problem

Radiologists at a large hospital system read hundreds of studies per day across multiple imaging modalities. Time-sensitive findings — intracranial hemorrhage on a non-contrast CT, pulmonary embolism on a CT angiogram, critical fractures — may be buried in a reading queue where the urgency is not immediately apparent. Radiology AI systems address this specific problem: they analyze incoming imaging studies, identify cases with potentially high-acuity findings, and re-prioritize the radiologist's worklist so that urgent studies are read first.

The problem is not AI accuracy — modern radiology AI achieves performance approaching or exceeding radiologist accuracy on specific tasks. The problem is integration: how does the AI output (a finding flag, a probability score, a bounding box overlay) enter the radiologist's actual workflow, at the point where it influences the read, without creating additional steps that radiologists reject as friction?

Why This Technology Exists

Medical imaging AI emerged from two converging developments: the availability of large labeled DICOM image datasets (driven by the digitization of radiology departments that followed PACS deployment in the 1990s and 2000s) and the maturation of convolutional neural network (CNN) architectures for image classification (ImageNet 2012, then specialized medical imaging research through 2015–2020).

The first wave of FDA-cleared radiology AI (2017–2020) focused on detection tasks: is this a finding of type X in this image? The second wave (2020–present) has moved toward quantification, triage, and workflow orchestration: how severe is this finding, which studies should be read first, and what measurements should appear in the radiologist's structured report?

Conceptual Explanation

DICOM (Digital Imaging and Communications in Medicine)

DICOM is the international standard for medical imaging data and communications. It defines:

  • Image storage format: A DICOM file (.dcm) contains both the image pixel data and a structured header (DICOM attributes) that stores patient demographics, study metadata, acquisition parameters, and equipment information.
  • Network protocol: DICOM defines services for sending, receiving, querying, and retrieving images between PACS, modalities, and viewing workstations.
  • Structured reporting: DICOM Structured Reports (SR) store coded measurements, observations, and annotations in a machine-readable format attached to an imaging study.

For AI integration, the relevant DICOM services are:

  • C-STORE: Sending a DICOM image from one system to another (used when PACS sends a study to an AI system for analysis)
  • C-FIND / C-MOVE / C-GET: Querying and retrieving specific studies from a PACS
  • DICOMweb (WADO-RS, STOW-RS, QIDO-RS): RESTful equivalents of the traditional DICOM network services, used by modern AI platforms that prefer HTTP over the traditional DICOM protocol stack

PACS (Picture Archiving and Communication System)

PACS is the system that stores, retrieves, and displays medical images. It is the central hub of the radiology workflow: images flow from imaging modalities (CT scanners, MRI machines, X-ray systems) to PACS, and from PACS to radiologist workstations for reading. For AI integration, PACS is the source of imaging studies and the destination for AI-enhanced worklist management.

Core Architecture

Components

DICOM Router / AI Orchestrator

The DICOM router is the component that receives incoming studies from PACS, routes them to the appropriate AI analysis service(s) based on modality and body part, and manages the return of AI outputs back to PACS and the worklist. For organizations deploying multiple AI products, the router is the orchestration layer: CT chest studies go to the pulmonary embolism AI; CT head studies go to the intracranial hemorrhage AI; chest X-rays go to multiple AIs simultaneously (pneumonia, pneumothorax, cardiomegaly).

AI Analysis Service

Each AI analysis service receives DICOM images, performs inference (typically on GPU), and returns results in DICOM-native formats:

  • DICOM Structured Report (SR): Coded findings, measurements, and probabilities in machine-readable format
  • DICOM Segmentation Object: Pixel-level masks identifying regions of interest (e.g., tumor boundary)
  • DICOM Presentation State: Annotations and overlays that appear on the radiologist's workstation when the study is opened
  • Probability score + worklist flag: The AI's confidence that a critical finding is present, used to re-prioritize the worklist

Worklist Integration

The radiology worklist determines which studies a radiologist reads and in what order. AI integration with the worklist re-prioritizes studies where AI identifies potentially critical findings to STAT priority. The specific mechanism depends on the RIS (Radiology Information System) and PACS:

  • Some PACS support direct worklist modification via HL7 messaging
  • Some AI vendors maintain a separate AI-prioritized worklist that complements (not replaces) the standard RIS worklist

FDA Regulatory Status

FDA clearance is the primary evaluation criterion for radiology AI products that a healthcare organization deploys in a clinical workflow where the AI output influences (even indirectly) diagnostic decisions. The deploying organization must confirm:

  1. The product has 510(k) clearance for the intended use (specific indication + modality + patient population)
  2. The intended use at the healthcare organization matches the cleared indication
  3. The cleared device description matches the software version being deployed

Cleared indications are highly specific: a 510(k) clearance for "adult chest CT for pulmonary embolism detection" does not cover pediatric patients or other modalities. Deploying the same software outside its cleared indication creates regulatory and liability exposure.

Implementation Patterns

DICOM Study Routing Configuration

python
# Educational Example — DICOM Study Router Configuration
# Illustrates how to configure routing rules for multi-AI radiology workflow
# Not a production DICOM router implementation

from dataclasses import dataclass
from typing import Optional


@dataclass
class AIRoutingRule:
    """
    A routing rule that determines when a study should be sent to an AI service.
    Based on DICOM metadata attributes.
    """
    rule_id: str
    description: str
    modality: str                      # DICOM (0008,0060) — e.g., "CT", "CR", "MR"
    body_part: Optional[str]           # DICOM (0018,0015) — e.g., "CHEST", "HEAD"
    study_description_contains: Optional[str]  # Substring match on study description
    ai_service_endpoint: str           # URL of the AI analysis service
    priority: str                      # "STAT" | "ROUTINE"
    cleared_indication: str            # FDA cleared indication for audit trail


RADIOLOGY_AI_ROUTING_RULES: list[AIRoutingRule] = [
    AIRoutingRule(
        rule_id="ct-head-ich",
        description="CT Head — Intracranial Hemorrhage Detection",
        modality="CT",
        body_part="HEAD",
        study_description_contains=None,
        ai_service_endpoint="https://ai.internal/radiology/ich-detection",
        priority="STAT",
        cleared_indication=(
            "Aid in the detection of intracranial hemorrhage in adult patients "
            "on non-contrast CT of the head. 510(k) K213174 equivalent."
        ),
    ),
    AIRoutingRule(
        rule_id="cta-pe",
        description="CT Angiography — Pulmonary Embolism Detection",
        modality="CT",
        body_part="CHEST",
        study_description_contains="ANGIOGRAPHY",
        ai_service_endpoint="https://ai.internal/radiology/pe-detection",
        priority="STAT",
        cleared_indication=(
            "Aid in the detection of pulmonary embolism in adult patients "
            "on CT pulmonary angiography. Verify current FDA clearance status."
        ),
    ),
    AIRoutingRule(
        rule_id="cr-chest-xr",
        description="Chest X-Ray — Critical Finding Triage",
        modality="CR",
        body_part="CHEST",
        study_description_contains=None,
        ai_service_endpoint="https://ai.internal/radiology/chest-xr-triage",
        priority="ROUTINE",
        cleared_indication=(
            "Aid in the prioritization of chest radiographs for radiologist review. "
            "Verify current FDA clearance status with vendor."
        ),
    ),
]


def route_study(
    modality: str,
    body_part: str,
    study_description: str,
    routing_rules: list[AIRoutingRule],
) -> list[AIRoutingRule]:
    """
    Determine which AI services should receive a given study based on routing rules.
    Returns all matching rules (a study may be sent to multiple AI services).
    """
    matched = []
    for rule in routing_rules:
        if rule.modality != modality:
            continue
        if rule.body_part and rule.body_part not in body_part.upper():
            continue
        if (
            rule.study_description_contains
            and rule.study_description_contains.upper() not in study_description.upper()
        ):
            continue
        matched.append(rule)
    return matched

Enterprise Considerations

Multi-Vendor AI Management: A radiology department deploying AI for 5–6 different indications will have AI products from 3–5 different vendors. Managing DICOM routing, version updates, performance monitoring, and regulatory compliance for multiple vendors simultaneously requires a radiology AI orchestration layer — a system that abstracts the individual AI product integrations behind a unified routing and monitoring interface.

Radiologist Trust Calibration: Radiologists who are new to AI-assisted workflows exhibit one of two problematic trust states: automation bias (accepting AI findings without independent assessment) or automation distrust (ignoring AI outputs entirely). Both states negate the clinical value of the AI. The change management program for radiology AI must specifically address trust calibration: when to use AI-identified findings as a starting point for independent assessment, and when to proceed independently.

AI Model Monitoring for Imaging Systems: Imaging AI models can experience distributional shift when scanner hardware is replaced, imaging protocols change, or patient demographics shift. A CT AI model trained on images acquired on one scanner type may degrade in performance on images acquired on a newer scanner from the same vendor with a different reconstruction algorithm. Radiology AI deployments require periodic performance monitoring against institution-specific ground truth, not just reliance on the vendor's published performance metrics.

Security Considerations

  • DICOM images contain PHI in the DICOM header (patient name, MRN, date of birth, study date) — any system that stores, transmits, or processes DICOM images is a PHI data store
  • Cloud-based AI analysis services that receive DICOM images require HIPAA BAA coverage
  • DICOM network traffic should be routed within the hospital's private network; DICOM studies should not be transmitted over public internet without end-to-end encryption
  • De-identification before sending to AI service: if the AI service does not require patient identifiers for its analysis function (most imaging AI does not), de-identify the DICOM header before transmission to the vendor

Healthcare Example

⊕ Healthcare Example

Educational Example — Illustrative Workflow. Not intended for clinical decision making.

The Reference Healthcare Organization deploys an intracranial hemorrhage (ICH) AI product (cleared as 510(k) SaMD) integrated into its emergency radiology workflow:

Integration architecture:

  • CT scanner sends completed head CT studies to PACS via C-STORE (unchanged from prior workflow)
  • PACS DICOM router forwards all CT head studies to the ICH AI service via DICOMweb STOW-RS simultaneously with standard archiving
  • ICH AI service analyzes the study in < 60 seconds and returns a DICOM SR with: ICH probability score, finding location, and estimated volume
  • If probability score > 0.85, the AI service sends an HL7 v2 message to the RIS to mark the study as STAT, re-prioritizing it in the radiologist worklist
  • The radiologist opens the STAT study; the AI overlay appears automatically in the PACS viewer showing the flagged region
  • The radiologist reads the study with the AI overlay visible, makes an independent assessment, and dictates the report

Performance outcomes (illustrative):

  • Time from scan completion to radiologist notification for positive ICH cases: reduced from 18 minutes (queue-dependent) to < 5 minutes (AI-triggered STAT)
  • Radiologist override rate: 12% (AI flagged as positive; radiologist confirmed negative after independent assessment — within expected specificity range for the cleared device)
  • No missed ICH cases in 6 months post-deployment (tracked via adverse event reporting and outcome audit)

Common Mistakes

Deploying Outside Cleared Indication. Using a 510(k)-cleared ICH detection AI on pediatric patients when the clearance covers adults only, or using it on MRI when cleared for CT, creates regulatory and liability exposure. Verify the cleared indication against the intended use before procurement and again before deployment.

DICOM Header PHI Exposure. Organizations that send DICOM studies to cloud AI services without DICOM header de-identification transmit patient identifiers (name, MRN, date of birth) to the AI vendor. Confirm whether the AI service requires patient identity, and de-identify the DICOM header if it does not.

Assuming Radiologist Will See the AI Overlay. AI overlays that require a separate workflow step to view (clicking a separate AI review button, opening a separate viewer) are rarely used. Design AI output delivery so that the AI overlay is visible in the radiologist's default study view without additional steps.

Best Practices

  • Verify 510(k) clearance status and cleared indication before procurement and before deployment
  • Confirm BAA coverage for any cloud-based AI service that receives DICOM images containing PHI
  • Deliver AI findings in the radiologist's standard PACS viewer — overlays that require workflow changes are not adopted
  • Monitor AI performance with institution-specific ground truth quarterly — vendor-published performance metrics may not generalize to your scanner hardware and patient population
  • Govern radiology AI through the same Model Review Board as other clinical AI use cases

Trade-offs

AI Integration Point Radiologist Workflow Impact Clinical Utility Implementation Complexity
Worklist prioritization only Minimal Medium Low
Worklist + passive overlay Low High Medium
Worklist + active overlay + structured report Medium Highest High
Autonomous read (no radiologist) Radical change Highest (if validated) Very high + regulatory

Interview Questions

Q: A radiology department wants to deploy AI that automatically flags intracranial hemorrhage on CT head studies. What regulatory and integration questions must be answered before deployment?

Category: Architecture / Regulatory Difficulty: Senior Role: AI Architect / Healthcare AI Consultant

Answer Framework:

Regulatory questions first: What is the product's FDA clearance status? Is there a 510(k) clearance for adult head CT ICH detection? What is the cleared indication — does it match the patient population (adult vs. pediatric)? Does the cleared indication cover the use as "triage" (flagging for radiologist review) or as "detection" (the AI identifies ICH)? These are different intended uses with different evidentiary requirements.

If the product is 510(k)-cleared for the intended use: does the healthcare organization need to do anything for regulatory compliance? In most cases, no — the 510(k) is held by the vendor, not the deploying organization. But the deploying organization must ensure the deployed version matches the cleared device description, and must have a process for applying FDA-required software updates in a timely manner.

Integration questions: How does the AI receive studies — C-STORE from PACS, or DICOMweb? How does the AI return its output — DICOM SR, worklist message, or overlay? Does the radiology workflow allow automatic re-prioritization of studies, or does the radiologist control the worklist? How long does AI inference take — is the latency acceptable for emergency use (under 60 seconds)?

HIPAA question: Does the AI service receive DICOM images in the cloud? If so, is a BAA in place? Does the AI require patient identifiers from the DICOM header, or can it operate on de-identified images?

Key Points to Hit:

  • 510(k) clearance verification, including patient population and intended use scope
  • Vendor holds the clearance; deploying org must deploy within the cleared indication
  • DICOM integration mechanism and AI output format
  • Worklist modification mechanism and radiologist workflow impact
  • HIPAA BAA if cloud-based

Key Takeaways

  • Radiology AI operates on DICOM images, not text — integration with PACS and the radiology worklist, not the EHR, is the primary engineering challenge
  • Almost all clinical radiology AI requires FDA 510(k) clearance as SaMD — verify clearance and match the cleared indication to the intended use before procurement
  • The DICOM router is the orchestration layer for multi-AI radiology workflows — it routes incoming studies to the appropriate AI services based on modality and indication
  • Deliver AI output in the radiologist's standard PACS viewer without additional workflow steps — overlays that require extra clicks are not adopted
  • Monitor performance with institution-specific ground truth — vendor performance metrics on public datasets may not reflect your scanner hardware and patient population

Glossary

DICOM: Digital Imaging and Communications in Medicine. The international standard for medical imaging data format, storage, and network communication.

PACS: Picture Archiving and Communication System. The system that stores, retrieves, and displays medical images in a radiology department.

510(k): An FDA pre-market notification that demonstrates a new device is substantially equivalent to a legally marketed predicate device. The most common regulatory pathway for radiology AI.

DICOM Structured Report (SR): A DICOM object that stores coded clinical observations, measurements, and findings in machine-readable format, attached to a study.

DICOMweb: RESTful web services (WADO-RS, STOW-RS, QIDO-RS) for accessing DICOM images using HTTP.

RIS (Radiology Information System): The information system that manages radiology department operations including scheduling, orders, worklist management, and reporting.

Further Reading