In a groundbreaking study published in Insights into Imaging, Savardi et al. (2025) explore the dual-edged impact of artificial intelligence on radiology training. As hospitals faced unprecedented challenges during the COVID-19 pandemic, integrating AI into diagnostic workflows became not only a necessity for patient care but also a transformative educational tool for radiology residents. This article delves into the study’s methodology, key findings, and the broader implications for medical education.
A New Frontier in Radiology Training
The study tackles an important question: Can AI support lead to upskilling—or does it risk deskilling—the next generation of radiologists? By incorporating an AI tool known as BS-Net into routine clinical practice, the research evaluates its impact on residents’ ability to assess lung compromise in COVID-19 patients via chest X-rays (CXRs). The BS-Net tool, integrated into the hospital’s RIS/PACS interface, demonstrated high performance in scoring lung severity, even outperforming the average radiologist in some metrics.
Methodology: Three Training Scenarios
Savardi et al. designed an experiment involving eight radiology residents who interpreted 150 CXRs under three distinct conditions:
- No-AI: Residents evaluated images without any AI assistance.
- On-demand AI: The AI support was available for residents to access voluntarily.
- Integrated AI: AI outputs, including confidence scores and explainability maps, were simultaneously displayed alongside the images.
By comparing performance across these scenarios, the study quantified the impact of AI on scoring accuracy, inter-rater agreement, and residents’ ability to manage AI errors.
Key Findings: Enhanced Accuracy and Consistency
The results of the study were striking:
- Increased AI Utilization: Residents requested AI support in 70% of cases, indicating a high perceived value in having an automated second opinion.
- Reduced Scoring Errors: Both on-demand and integrated AI scenarios significantly lowered the average scoring error compared to the no-AI scenario.
- Improved Agreement: Inter-rater agreement among residents improved by 22% when AI support was available, as measured by the intraclass correlation coefficient (ICC).
- Resilience to AI Errors: Even when the AI provided erroneous predictions beyond an acceptable threshold, residents demonstrated the ability to critically assess and adjust their scores—especially those with more advanced training.
These findings underscore that AI, when properly integrated, can enhance diagnostic precision without compromising the development of critical clinical skills.
Upskilling vs. Deskilling: A Balanced Approach
One major concern in the era of AI is the risk of deskilling—where reliance on automated tools might erode essential human expertise. However, the study’s outcomes suggest a more nuanced picture:
- Upskilling Effects: The AI tool served as a valuable training aid, offering immediate feedback and aiding in the standardization of assessments through explainability maps and confidence indicators.
- Deskilling Safeguards: Importantly, residents did not become overly dependent on the tool. Their resilience in the face of AI errors indicates that they maintained a critical approach to AI outputs, a crucial factor in preventing deskilling.
The study highlights the importance of collaborative interaction with AI, where the tool supports, rather than replaces, the radiologist’s judgment.
Implications for Radiology Residency Programs
The integration of AI into radiology training programs can offer several benefits:
- Enhanced Learning: Immediate, AI-generated feedback helps residents quickly identify and correct errors.
- Increased Consistency: Standardized scoring reduces inter-observer variability, leading to more reliable diagnostic assessments.
- Adaptability: Training under varied scenarios (no-AI, on-demand, and integrated AI) equips residents with the skills to leverage AI effectively while maintaining independent critical thinking.
By carefully balancing the benefits of AI support with strategies to mitigate the risk of deskilling, educational programs can prepare residents for a future where human expertise and machine efficiency work hand in hand.
Conclusion
Savardi et al.’s study provides compelling evidence that AI can be a powerful ally in radiology education. When integrated thoughtfully into clinical workflows, AI not only improves diagnostic accuracy and consistency but also enhances the training experience of residents—helping them develop a more robust and nuanced understanding of complex imaging tasks. These insights offer a promising roadmap for incorporating AI into medical education while safeguarding the critical skills that define expert clinical practice.
Reference: Savardi et al., “Upskilling or deskilling? Measurable role of an AI-supported training for radiology residents: a lesson from the pandemic,” Insights into Imaging, 2025