From the Inside: FAQ with Engineering

Over the past few decades, radiology has seen wave after wave of technological advancement: from film to digital imaging, from PACS to voice recognition. This pace of innovation is not new. What is new are the mounting and intensifying pressures on today’s health systems: ever-increasing imaging volumes, radiologist shortage, and growing demands for efficiency without sacrificing quality. To meet these challenges, radiologists and health systems are now turning to AI not as a replacement, but as a trusted partner to ease workloads, reduce burnout, and unlock new levels of clinical impact.

In this blog, our VP of Engineering, Ryan Hood, answers six of the most frequently asked engineering questions, sharing how Rad AI approaches AI product development, safety, transparency, and real-world adoption.

How does the engineering team ensure that AI is implemented as an assistive tool and not as a replacement for radiologists?

From the outset, Rad AI has been centered around assisting, not replacing, radiologists. While others focused on imaging automation, we focused on reducing fatigue and burnout by streamlining workflows. 

We have five radiologists on staff who provide us constant feedback, and through shadowing programs and close collaboration, we can focus on and prioritize feature enhancements that radiologists truly care about.

In the case of our Impressions product, the model is trained on the specific preferences of each radiologist. This allows us to personalize the Impressions to the radiologist’s unique preferences, which results in a high acceptance rate from the radiologists. As another example, our OmniBox feature of our Reporting product allows radiologists to dictate freely into the OmniBox, and our models map the text back to the relevant sections of the report template so that the radiologists do not need to manually hop around each section of the template to fill in the appropriate clinical data. This allows radiologists to focus on the highest value they can add: clinical insights.

AI is powerful and capable of many wonderful things, but we still have not reached the stage where a fully self-driving vehicle is authorized to drive on freeways and within cities. Driving a car is a reasonably straightforward task for an adult human. However, when you look at something like radiology, we are dealing with something very complex where nuance and high judgment impact the quality of care. That is why we always keep the radiologist at the forefront of everything we build. We would struggle to make things useful for them if we had less access to our internal radiologists and did not work in lockstep. 

What specific features or processes are in place to keep the radiologist “in the loop” and ensure final decisions remain human-led?

We've architected our entire system around a fundamental principle: AI generates or modifies draft reports only. It cannot finalize or sign them. This isn't just a technical limitation; it's a deliberate design choice that preserves the critical legal and clinical accountability that belongs exclusively with the radiologist.

Our visual transparency system displays all AI-generated content in a distinct font color throughout the report. This immediate visual distinction serves multiple purposes: it helps radiologists quickly identify which sections need their particular attention, maintains awareness of AI involvement, and creates a clear audit trail. Whether it's AI-generated impressions or voice-to-text transcription, radiologists always know precisely what originated from their clinical judgment versus AI assistance.

The multi-author workflow feature adds another layer of human oversight, enabling multiple radiologists to collaborate on complex cases, each able to review, edit, and sign reports. This is particularly valuable for challenging cases or teaching scenarios where senior radiologists review junior colleagues' work alongside AI suggestions.

Behind the scenes, our internal radiologists conduct near-daily quality monitoring, reviewing AI outputs and ensuring our assistance aligns with real-world needs. This continuous human oversight happens at both the individual report level and system-wide, with quantitative metrics and qualitative reviews feeding back into our improvement cycles.

This comprehensive approach means that while AI handles the repetitive and time-consuming aspects of report generation, every clinical decision, diagnosis, and signed report represents deliberate human judgment. The radiologist remains the ship's captain. We just help them navigate more efficiently.

How does the engineering team approach transparency in how AI arrives at its decisions?

We build transparency into our systems at multiple levels. As mentioned above, we display all AI-generated content in a different font color, making it immediately clear to radiologists what the AI has produced versus their own input. This visual distinction is critical for maintaining trust and enabling proper review.

For our newer features in development, we're building additional explainability layers that will surface the specific reasoning pathways and data lineage behind AI decisions. While our established products like Impressions have earned trust through years of consistent performance and radiologist validation, we recognize that novel AI capabilities will benefit from more granular transparency about their decision-making process. This graduated approach ensures we provide the right level of explanation for each use case: proven tools focus on performance transparency, while emerging features will include more detailed decision tracing.

How do you gather and incorporate radiologists' feedback about onboarding and day-to-day use of AI?

Feedback is at the heart of the user experience for our products, starting from onboarding. During initial rollout, we work closely with each practice, shadowing radiologists when necessary and addressing friction points in real time. This helps ensure the transition feels seamless and that the AI is adapting to their existing workflow, not the other way around.

Once in daily use, radiologists can provide feedback to us directly within the Impressions and Reporting app anytime. We make it fast and easy to offer feedback, yet we also allow for as many details as the users feel necessary to express their feedback. 

We offer categories that fit their themes so that we can quickly see if specific categories are receiving an outsized amount of feedback and then drill into the specifics from there to address the root cause. We review the submitted feedback in larger forums weekly to ensure we are up-to-date with trends we hear from the ground. 

We also regularly shadow radiologists to understand problems that are particularly difficult to troubleshoot or need more information. This also helps us understand their issues, needs, and wants in more detail. By receiving direct feedback and learning about each radiologist's specific habits and preferences, we can customize our model to them, providing a truly personalized experience. 

One of our core values is that “every radiologist matters," and radineers take that to heart. Things that may seem like minor issues from an engineer’s point of view, like a cursor jumping from one field to another or a misplaced space or comma, can result in a meaningful productivity loss for radiologists. That’s why we listen closely, act quickly, and make radiologists’ feedback central to development, creating a product that fits their needs today and keeps evolving as those needs change.

What strategies are in place to manage ethical challenges, such as accountability for errors or data privacy?

We've built a comprehensive framework addressing ethical challenges across multiple dimensions, with accountability and privacy as foundational pillars.

For accountability and error management, our AI systems are designed to augment, not replace, radiologist judgment. The AI cannot sign reports, and all AI-generated content appears in a different font color for transparent physician review. Our internal radiologists conduct near-daily quality monitoring, while feedback mechanisms help us quickly identify and address any performance issues.

Regarding data privacy, we undergo annual third-party HIPAA compliance auditing alongside SOC 2 Type II certification. We implement comprehensive security measures, including encryption, role-based access controls, audit logging, and secure data deletion protocols.

We employ a risk-based governance approach that prioritizes high-risk use cases for additional scrutiny, actively mitigates bias through diverse training datasets, and maintains transparent communication about our AI's capabilities and limitations. Every model deployment follows a rigorous checklist that considers patient safety, bias potential, and real-world impact before going live. This multi-layered approach ensures we harness AI's power to improve patient care while never compromising ethical responsibilities or patient safety.

What is your perspective on the future role of radiologists as AI becomes more integrated into imaging workflows?

It will be incredibly exciting to be a radiologist in the future. AI enhancements will remove many repetitive tasks and boilerplate that have defined the workflow for decades, allowing radiologists to focus their expertise on complex, intricate, and rare cases that require their cognitive bandwidth.

We've already seen this transformation in other fields. Consider agent assist features in call centers - what used to require agents to spend minutes searching internal knowledge bases, clicking through documentation, and piecing together information now takes seconds. AI provides highly relevant summaries and clear action steps instantly. The agents still apply their judgment and empathy to solve customer problems, but they're freed from the tedious information gathering that once consumed so much of their time.

Similarly, think about how generative AI has changed how we all search for information. Instead of clicking through dozens of links to find the correct answer, we get comprehensive summaries synthesizing multiple perspectives. The technology doesn't replace our thinking; it accelerates our access to information.

This same principle applies to radiology. AI preserves everything that makes radiologists essential: their medical expertise, pattern recognition, and clinical judgment, while removing the manual, repetitive tasks that have historically consumed so much time. Nothing is more taxing than frequent repetitive manual tasks, which, unfortunately, have been commonplace in traditional radiology workflows.

With these efficiency gains, radiologists will not only become more effective in their practice but also experience much higher job satisfaction. They'll spend less time on routine documentation and more time on the intellectually rewarding aspects of their profession—the cases that challenge them and the patient interactions that matter most.