Key Takeaways
- A Northwell Health study published in Radiology: Artificial Intelligence found that Aidoc‘s AI algorithm agreed with radiologist interpretations in nearly 98% of cases across more than 32,000 CTPA scans.
- AI-informed radiologists achieved 99.2% sensitivity for pulmonary embolism detection, higher than either AI or human interpretation alone, supporting a human-in-the-loop diagnostic model.
- Radiologists identified an additional 15% of confirmed positive pulmonary embolism cases that Aidoc’s AI missed, meaning enterprises deploying the technology still require full radiologist review rather than AI-only triage. A study of more than 32,000 CT pulmonary angiography scans has produced one of the most detailed real-world assessments yet of AI-assisted pulmonary embolism detection, and its most important finding is not about what the AI got right. Published in Radiology: Artificial Intelligence by researchers at Northwell Health, the study found that radiologists working alongside Aidoc’s algorithm achieved 99.2% sensitivity for PE detection, outperforming either approach alone, but radiologists still caught 15% of confirmed positive cases the AI had missed, a gap that carries direct consequences for how healthcare enterprises should deploy these tools.
How AI-Powered Systems and Radiologist Interpretation Compare
Deciding whether to adopt AI-powered diagnostic tools or rely on traditional radiologist workflows involves more than a performance comparison. Healthcare enterprises need to weigh enterprise use cases, cost, scalability and integration complexity side by side.
AI-Powered Pulmonary Embolism Detection
AI platforms for PE detection, including those from Aidoc, Viz.ai and Avicenna.AIare built primarily as triage and notification tools. Aidoc’s algorithm analyses CTPA scans and immediately flags suspected positive cases, allowing radiologists to prioritise the most urgent studies first. In high-volume settings where PE can be life-threatening, that prioritisation matters. Avicenna.AI’s CINA-iPE, which has received FDA clearance, extends this further by identifying incidental pulmonary embolisms during CT scans ordered for other conditions, catching findings that might otherwise be delayed. Viz.ai’s PE solution focuses on AI-powered alerts for suspected PE and right heart strain, with the goal of faster care team coordination.
Beyond detection, these platforms contribute to care coordination by automatically notifying radiologists, pulmonologists, cardiologists and vascular surgeons when a high-risk PE case is identified. The practical effect is compressing the time from scan to diagnosis from hours to minutes, which can meaningfully affect treatment outcomes.
On cost, AI solutions typically involve subscription or licensing fees, but the return case rests on efficiency: faster diagnosis, shorter hospital stays and better-optimised radiologist time. Aidoc’s enterprise model, for instance, is oriented around hospital-wide workflow integration rather than point-of-care use, which changes how the cost-benefit calculation looks for large health systems.
Scalability is where AI has the clearest structural advantage. An algorithm can process scans continuously around the clock without fatigue or capacity constraints. The Northwell Health study, which covered 32,501 CTPAs over 18 months, reflects what that looks like in practice. On integration, both Aidoc and Viz.ai have built their platforms to fit into existing Picture Archiving and Communication Systems (PACS) and Electronic Health Record (EHR) environments, with automated alerts delivered through mobile and web applications. Getting there still requires planning, IT support and staff training, but the tooling is designed for current clinical infrastructure, not a replacement of it.
Traditional Radiologist Interpretation
Radiologists bring something current narrow AI systems cannot replicate: the ability to integrate a patient’s full clinical context with image findings, identify incidental pathology beyond the primary question and exercise judgment in ambiguous cases. That capability is not a soft advantage. The Northwell Health study made it concrete: radiologists recovered 15% of confirmed PE cases that Aidoc’s algorithm had missed. That figure is the clearest argument against treating AI as a standalone diagnostic tool.
The cost of maintaining radiology departments is substantial, covering salaries, benefits, training and departmental overhead. The strategic question for enterprises is whether that investment is being directed toward the cases where human judgment is most valuable, or whether radiologists are spending significant time on high-volume, lower-complexity reads that AI could handle first. Scalability remains the sector’s persistent pressure point. The pool of trained radiologists is finite, demand for imaging continues to grow, and the resulting workload contributes to fatigue and burnout, particularly in under-resourced or remote settings. AI adoption is partly a response to that structural problem.
Traditional workflows are well-established within PACS and EHR systems, but speed in acute triage is where they are most vulnerable. Without AI-assisted flagging, the manual review queue can delay diagnosis of time-sensitive conditions like PE, particularly during off-hours or in high-volume periods. That is the gap AI is most directly addressing.
What the Northwell Data Actually Shows
The 97.8% agreement rate between Aidoc’s algorithm and radiologist interpretations across more than 32,000 scans is a strong performance benchmark, but the more operationally significant number is 99.2%: the sensitivity achieved when radiologists worked with AI assistance rather than without it. That improvement reflects the core value proposition for enterprise adoption, not AI replacing radiologists, but AI making radiologists faster and more thorough.
The 15% of confirmed positive PE cases caught only by radiologists after AI missed them is an equally important number. It establishes a hard floor: full radiologist review remains necessary. AI excels at rapid, high-volume triage and as a safety net for catching cases that might otherwise wait in a queue. It does not yet perform reliably enough to function as the terminal step in diagnosis. For enterprise decision-makers, that distinction shapes where AI fits in the workflow and what governance structures need to stay in place around it. For a broader look at how AI evaluation frameworks are evolving in high-stakes environments, the Mpathic mPACT benchmark research on AI safety evaluations is worth reviewing alongside this study.
Recommendations for Enterprise Adoption
The Northwell Health study points clearly toward a human-AI collaborative model rather than an either/or deployment decision. For enterprises building or refining their diagnostic imaging strategy, several practical implications follow from the evidence.
Deploy AI as a triage and augmentation layer, not a replacement for radiologist sign-off. The 99.2% sensitivity figure was achieved with radiologists in the loop; the 15% miss rate establishes why that loop must stay closed. Prioritise platforms with regulatory clearance and large-scale real-world validation rather than lab benchmarks alone. Aidoc, Viz.ai and Avicenna.AI all have documented clinical deployments, which matters for procurement decisions where evidence requirements are high.
Integration quality is as important as detection accuracy. An AI tool that flags suspected PE but cannot deliver that alert reliably into existing PACS and EHR workflows, or reach the right care team members in time, does not deliver its theoretical value. Evaluate platforms on their notification architecture and workflow fit, not just their algorithm performance. Staff training is not optional: radiologists who understand how an AI tool reasons, where its failure modes are and how to interpret its outputs will use it more effectively than those who treat it as a black box.
Finally, look beyond binary detection. Tools offering quantitative outputs such as right ventricle to left ventricle ratio for PE risk stratification, or incidental finding detection across scan types, extend the diagnostic value of a single platform investment. The evidence base for AI in radiology is maturing quickly; the enterprises that build governance structures now will be better positioned to adopt more capable tools as they become available. For more coverage of AI research and breakthroughs, visit our AI Research section.
Originally published at https://autonainews.com/aidoc-ai-matches-radiologists-lifts-pe-detection-to-992/
Top comments (0)