AI can raise radiology accuracy and throughput, but benefits are conditional on real-world integration—poorly designed deployments risk automation bias, deskilling and workflow disruption that can erode clinical and economic gains.

Main Finding

Human–AI collaboration in radiology is most effective when designed around diagnostic complementarity and physician-in-the-loop workflows: AI augments radiologists by handling high-volume, repetitive, and pattern-recognition tasks while humans retain contextual judgment, oversight, and final responsibility. Properly structured collaborations can improve diagnostic accuracy, timeliness, and efficiency, but they require governance, continuous monitoring, explainability, and workforce adaptation to avoid harms (automation bias, algorithmic aversion, deskilling).

Key Points

• Diagnostic complementarity and human–AI symbiosis: Combining imperfect, partially non-overlapping error sets of humans and AI can yield team performance superior to either alone; clear role separation and interaction design are critical.
• Spectrum of autonomy: Interaction models range from concurrent decision support and second-reader workflows to triage/prioritization and high-confidence auto-finalization; fully autonomous clinical reporting remains conceptual and not currently viable.
• Empirical examples:
  • Mammography simulations and prospective deployments show AI can reduce human workload (e.g., supporting-reader workflows cut second reads) and modestly increase cancer detection (additional ~0.7–1.6 cancers per 1,000 screens in one prospective study).
  • A multi-reader prostate MRI study (61 readers, 360 exams) found AUC improved from 0.882 to 0.916 with AI assistance (sensitivity and specificity gains).
  • AI-based triage/prioritization studies report large reductions in turnaround times for critical findings (examples: mean reporting delays reduced from ~11.2 to 2.7 days in a simulation; CT pulmonary angiography turnaround from 59.9 to 47.6 minutes in deployment).
• Risks and cognitive impacts: automation bias (overreliance), algorithmic aversion (underuse), cognitive offloading leading to deskilling, context-dependent workload changes, and disproportionate vulnerabilities for trainees.
• Design and implementation requirements: configurable, workflow-embedded tools; explainability and usable outputs; explicit manual oversight/arbitration mechanisms; closed-loop learning/active learning for continuous model refinement; AI literacy and co-design with radiology teams.
• Governance and safety: need for formal liability frameworks, continuous post-market surveillance for performance drift, explainability standards, and prospective multi-institutional evaluation focused on team performance, equity, and long-term training outcomes.
• Emerging tech: vision–language models and adaptive learning loops (active learning) offer draft-reporting and summarization potential but need robust evaluation and guardrails.

Data & Methods

• Paper type: narrative (non-systematic) review based on targeted topic-driven literature searches, expert knowledge, and reference chaining.
• Evidence synthesized: conceptual models, qualitative field studies, simulation and retrospective reader studies, prospective clinical implementations, and multi-reader diagnostic studies. Representative empirical items described include:
  • Large-scale mammography simulation (Ng et al.) using >280,000 exams across sites and vendors assessing “AI as supporting reader”.
  • Prospective mammography implementation where AI as an additional reader increased cancer detection modestly with minimal recall increase.
  • Multi-center prostate MRI diagnostic study (Prostate Imaging–Cancer AI Consortium): 61 readers across 53 centers, 360 examinations—AUC improvement with AI assistance.
  • Reader-study simulations showing human–AI team gains (e.g., Lee et al., 30 radiologists evaluating chest radiographs).
  • Triage/prioritization simulations and deployments showing substantial turnaround time reductions for urgent findings.
  • Qualitative cross-regional field study (Zając et al.) across Denmark and Kenya on radiologists’ visions for triage and workload distribution.
• Limitations acknowledged: narrative review methodology (not systematic), dependence on selected studies that vary in design/quality, and the relative scarcity of prospective, multi-site randomized assessments of human–AI team performance and long-term workforce effects.

Implications for AI Economics

• Labor augmentation vs substitution: The paper supports an augmentative model where AI increases radiologist productivity and throughput rather than wholesale substitution. Economic outcomes depend on the degree of task automation, local practice patterns, and regulatory constraints.
• Productivity and capacity effects:
  • Short-term: AI triage and automation of manual tasks can raise throughput, shorten turnaround times, and reduce backlog—potentially increasing revenue per radiologist or enabling reallocation of radiologist time to higher-value activities.
  • Medium/long-term: Persistent efficiency gains may compress per-case prices (if supply expands) or enable greater volume of imaging (demand elasticity), affecting imaging service revenues.
• Human capital and training costs:
  • Risk of deskilling and the need for retraining entail investment in AI literacy, new curricula, and supervised practice. These are recurrent costs for hospitals, training programs, and payers.
  • Trainee vulnerability implies potential long-term impacts on the supply and skill composition of the radiology workforce, with implications for wage dynamics and credentialing requirements.
• Implementation and maintenance costs:
  • Beyond license fees, institutions face integration (PACS/EHR), customization, governance, continuous monitoring, data labeling/active learning, and liability/insurance costs—raising the total cost of ownership and favoring larger institutions with scale to amortize these expenses.
  • Ongoing model maintenance (drift monitoring, retraining) creates recurring operational expenditures and staffing demands (data engineers, AI safety officers).
• Market structure and competition:
  • Economies of scale and network effects (better models with more data/labels) may concentrate market power among a few large vendors, increasing risks of vendor lock-in and raising bargaining leverage for incumbents.
  • Differentiation will favor vendors who provide robust workflow integration, physician-in-the-loop tooling, explainability, and regulatory compliance.
• Reimbursement and regulatory uncertainty:
  • Economic viability depends on reimbursement models—whether payers reimburse AI-assisted reads, triage services, or only radiologist interpretation. Clear coding/reimbursement pathways (e.g., add-on billing for AI-assisted workflows) will shape adoption speed.
  • Liability allocation (who is responsible for errors in physician-in-the-loop vs autonomous systems) influences malpractice insurance and risk premiums—uncertainty can slow adoption and increase costs.
• Value and downstream healthcare economics:
  • Improved detection (e.g., additional cancers detected) can raise downstream treatment costs but also potentially improve outcomes; cost-effectiveness depends on true-positive vs false-positive tradeoffs, overdiagnosis risks, and downstream care pathways.
  • Efficiency gains could enable reallocation of radiologist time to value-enhancing activities (multidisciplinary care, interventional procedures), changing revenue mixes across departments.
• Distributional and equity concerns:
  • Resource-rich centers are more able to absorb implementation and governance costs, potentially widening performance and outcome gaps between institutions and regions—raising equity issues in access to advanced imaging diagnostics.
• Investment implications:
  • Investors should value firms that prioritize human-in-the-loop designs, active-learning features that lower labeling costs, strong integration/API capabilities, explainability, and compliance-ready offerings.
  • Prospective evaluation data (multi-site trials demonstrating team-level clinical and economic benefits) will materially affect adoption risk premiums and valuations.
• Metrics and policy needs:
  • Policymakers and payers should incentivize prospective, team-focused evaluations (not just model accuracy), fund training and surveillance infrastructure, and clarify reimbursement and liability frameworks to align economic incentives with safe adoption.

Summary recommendation for economists and decision-makers: evaluate AI investments not solely on model performance but on total cost of ownership (integration, governance, retraining), measurable improvements in team-level productivity/outcomes, and distributional effects across institutions and workforce cohorts. Prioritize procurement of systems designed for physician-in-the-loop use, with transparent monitoring and active-learning capabilities that reduce long-run labeling costs and performance drift.