EXPLORING CLINICIANS PERSPECTIVES TOWARDS AI-RADIOLOGY & ITS CLINICAL ADOPTION: A QUALITATIVE STUDY FROM PAKISTAN

This qualitative study of 13 clinicians in Karachi, Pakistan, reveals that while AI in radiology is viewed as a promising assistant for specific tasks like screening and workload reduction, its successful adoption in this low-resource setting depends on addressing infrastructure gaps, ensuring data privacy, providing targeted training, and maintaining human oversight through strategic, staged implementation.

The Gist

Imagine you are a doctor in a busy hospital in Karachi, Pakistan. You have a stack of X-rays, CT scans, and MRIs to read. Your eyes are tired, your coffee is cold, and you know that missing even one tiny detail could change a patient's life.

Now, imagine a new tool has arrived: Artificial Intelligence (AI). It's like a super-smart, tireless intern that can look at these scans in seconds and point out potential problems.

This research paper is essentially a group of 13 doctors (radiologists and specialists) sitting down for a coffee chat to answer one big question: "Are we ready to let this AI intern help us, or is it going to cause more trouble than it's worth?"

Here is the story of their conversation, broken down into simple concepts and analogies.

1. The Big Picture: "Optimism with a Safety Net"

The doctors aren't scared of AI, but they aren't ready to hand over the keys to the car either. Their attitude is best described as "Realistic Optimism."

• The Analogy: Think of AI not as a replacement driver, but as a very advanced GPS.
  • The GPS (AI) can tell you the fastest route, warn you about traffic jams (tumors or fractures), and help you avoid getting lost.
  • But the doctor is still the driver. The doctor must keep their hands on the wheel, make the final decision, and take responsibility if they hit a pothole.
  • The Consensus: "AI is a great assistant, but it cannot replace the doctor's brain."

2. The Good Stuff: Why They Want the GPS

The doctors see huge benefits if the technology works right. They imagine a future where:

• The "Super-Speed" Triage: Imagine a waiting room with 100 patients. The AI acts like a traffic cop, instantly flagging the 5 patients who need immediate attention so the doctors don't have to guess who to see first.
• The "Second Pair of Eyes": When a doctor is exhausted after a 12-hour shift, the AI acts like a safety net, double-checking the work to make sure nothing subtle was missed.
• The "Remote Explorer": In rural areas where there are no expert doctors, the AI could act as a bridge, helping local clinics screen patients and send only the tricky cases to the big city.

3. The Bumps in the Road: Why They Are Hesitant

Despite the benefits, there are major hurdles. The doctors point out that the "GPS" is currently broken in several ways:

• The "Black Box" Problem: Sometimes the AI gives an answer, but it can't explain why. It's like a GPS saying, "Turn left," but refusing to tell you if there's a bridge out. Doctors need to trust the logic, not just the result.
• The "Leaky Bucket" (Privacy): Doctors are terrified that patient data might get stolen. They worry that putting sensitive medical images on the internet is like leaving your house keys under the doormat. They need strict locks (encryption) and rules.
• The "Old Car" Problem (Infrastructure): Karachi has great hospitals, but the internet connection is often shaky, and the computers are old. Trying to run high-tech AI on slow internet is like trying to race a Ferrari on a dirt road; it just won't work smoothly.
• The "Training Gap": Most doctors learned medicine before AI existed. They feel like they were handed a spaceship manual without ever being taught how to fly. They need schools and workshops to learn how to use these tools.

4. The Solution: How to Make It Work

The doctors didn't just complain; they offered a blueprint for the future. They suggest that for AI to succeed in Pakistan, three things must happen:

1. The "License Plate" Check (Validation): Before any AI tool is allowed in a hospital, it needs to be tested locally. Just because it worked in the US or Europe doesn't mean it works for Pakistani patients. It needs a local safety inspection.
2. The "Driver's Ed" (Education): Medical schools need to start teaching AI. Future doctors need to learn how to talk to the AI, how to spot when it's wrong, and how to use it as a tool, not a crutch.
3. The "Rulebook" (Policy): The government needs to write clear laws. Who is responsible if the AI makes a mistake? How is data protected? We need a rulebook so everyone knows the rules of the road.

The Final Takeaway

The paper concludes that AI in radiology is not a sci-fi movie where robots take over. Instead, it's a partnership.

If we fix the internet, teach the doctors, and build strong safety rules, AI can be the ultimate co-pilot. It won't replace the doctor, but it will help them fly higher, safer, and faster, ensuring that every patient gets the best care possible.

In short: The doctors are ready to embrace the future, but they want to make sure the plane is safe, the pilot is trained, and the map is accurate before they take off.

Technical Summary

1. Problem Statement

Despite the global proliferation of Artificial Intelligence (AI) in healthcare, there is a significant evidence gap regarding its adoption in Low- and Middle-Income Countries (LMICs), specifically Pakistan. Existing literature predominantly focuses on:

• Technical performance metrics of AI algorithms.
• Implementation in High-Income Countries (HICs).
• The Gap: There is a lack of qualitative research exploring the attitudes, beliefs, and practical barriers faced by clinicians in LMIC settings. Without understanding these human factors, AI implementation risks low uptake, improper utilization, and ineffective resource distribution, potentially compromising health system objectives.

2. Methodology

The study employed a qualitative descriptive exploratory design grounded in a constructivist paradigm (relativist ontology).

• Study Setting: Tertiary care hospitals in Karachi, Pakistan (representing a mixed-resource setting with public, private, and military facilities).
• Participants: 13 clinicians and radiologists (Consultants and Specialists) from diverse specialties including Medicine, Surgery, Gynecology, Radiology, and Pediatrics.
• Data Collection: Semi-structured, face-to-face interviews conducted between May and August 2025. Interviews were audio-recorded, transcribed, and anonymized.
• Analytical Frameworks:
  • Technology Acceptance Model (TAM): Used for deductive content analysis to categorize data based on Perceived Usefulness and Perceived Ease of Use.
  • Problem-Driven Iterative Adaptation (PDIA): Used to interpret findings and formulate practical policy pathways and implementation strategies.
• Data Analysis: Thematic analysis performed using NVivo 14 software. Rigor was ensured through triangulation (data, investigator, and theoretical), reflexivity, and adherence to the COREQ checklist.

3. Key Contributions

This study provides the first locally grounded evidence from Pakistan regarding clinician perspectives on AI in radiology. Its primary contributions include:

• Contextualizing AI in LMICs: Moving beyond technical feasibility to address the socio-technical realities of a resource-constrained environment (e.g., connectivity issues, infrastructure gaps).
• Bridging Theory and Practice: Integrating TAM (user acceptance theory) with PDIA (policy implementation) to offer a roadmap for adoption that is both theoretically sound and practically actionable.
• Identifying Specific Barriers: Highlighting unique challenges in Pakistan, such as generational resistance, lack of local validation, and specific data privacy concerns in a developing regulatory landscape.

4. Key Results & Findings

The analysis yielded four overarching themes:

A. Attitudes Towards Implementation

• Realistic Optimism: Clinicians view AI as a "second reader" or assistant, not a replacement.
• Accountability: There is a unanimous consensus that clinical judgment and ultimate accountability cannot be delegated to AI.
• Trust Deficit: Significant concerns exist regarding data privacy, cybersecurity, and the potential for AI to generate "fake" or hallucinated data.

B. Perceived Ease of Use (Barriers vs. Drivers)

• Barriers:
  • Infrastructure: Poor internet connectivity and lack of integration with existing Radiology Information Systems (RIS/PACS).
  • Training: Absence of structured training programs and upskilling opportunities for existing staff.
  • Generational Resistance: Older clinicians exhibit higher resistance due to a lack of digital literacy.
• Drivers:
  • Potential for workload reduction and burnout prevention.
  • Support for mass screening and triage acceleration.

C. Perceived Usefulness

• High Value Areas: AI is perceived as most useful for specific bottlenecks: quick screening, quantitative measurements, remote-area reporting, and decision support for trainees.
• Critical Limitations: Concerns regarding algorithm errors, lack of generalizability to local populations, and the inability of AI to replace contextual clinical reasoning.

D. Recommendations and Policy Practice

• Validation: A strict local vetting and pilot approval process is required before clinical deployment.
• Education: Integration of AI literacy into medical curricula and continuous professional development (workshops).
• Governance: Establishment of national data protection policies, ethical guidelines, and clear liability frameworks (defining responsibility for AI errors).

5. Significance and Implications

• Policy Formulation: The study provides a blueprint for Pakistani policymakers to create regulations that ensure AI tools are validated, secure, and ethically deployed. It emphasizes the need for a "human-in-the-loop" approach.
• Strategic Implementation: It suggests that successful adoption requires a staged approach: first addressing infrastructure and trust, followed by training, and finally integration into routine workflows.
• Global Relevance: As a case study from an LMIC, it offers transferable insights for other developing nations facing similar resource constraints and the need to balance technological advancement with human oversight.
• Future Outlook: The authors conclude that AI will not replace radiologists but will serve as a collaborative partner. However, realizing this potential requires a multi-dimensional strategy involving rigorous local testing, infrastructure investment, and the establishment of trust through transparency and evidence.

Limitations: The study is limited by its small sample size (n=13) and geographic restriction to Karachi, which limits generalizability to rural or secondary care settings in Pakistan. However, the depth of qualitative insight offers high value for hypothesis generation and policy direction.