Becoming AI’s best partner: A new reality in neuroradiology (VIOLA-AI)

Mario Mahecha, Renata Tarraf Fernandes, Santiago Guzman, Santiago Aristizábal, Pablo Riascos

---

Recently one member of our group shared a story about his grandfather, who had suffered a small brain hemorrhage. The first radiologist who read the CT scan did not see it but later another radiologist reviewed the same images and spotted the bleed. This sparked a question in our discussion: “Is detection of these small hemorrhages completely standardized with tools, or does it still depend mainly on expertise?”

That question led us to look deeper into the literature. We found that, indeed, small hemorrhages can sometimes be missed, not only by AI systems but also by radiologists, especially in high-pressure emergency settings. But we also came across something encouraging: a recent project in Norway where an AI tool, VIOLA-AI, was tested in real-world hospital practice (different from most AI applications that are tested today, which usually remain in research or pilot phases and are not actively learning in real clinical settings), designed specifically to learn from neuroradiologists over time.

What is VIOLA-AI?

VIOLA-AI (Voxels Intersecting along Orthogonal-Levels of Attention) is a deep learning system built for detecting intracranial hemorrhage (ICH) on CT scans. It integrates directly into the radiologist’s normal viewer, giving results in just a few minutes:

• Red highlights when a hemorrhage is detected, along with an estimated size.

• A green mark if no hemorrhage is found.

  
  

This allows it to act as a quick second opinion while the radiologist prepares the final report.

How it learns from radiologists

What makes VIOLA-AI (Voxels Intersecting along Orthogonal-Levels of Attention) different is that it runs on the NeoMedSys platform, which allows direct feedback. If the AI misses a small bleed or incorrectly flags a normal structure, the neuroradiologist can correct it. These corrections are then fed back into the system, and the model is retrained regularly.

This means the AI is not a fixed tool, it actually adapts and improves with the hospital’s own data, population and the expertise of its radiologists.

The results

The Norway trial showed how effective this approach can be:

• Sensitivity (catching true bleeds) improved from ~79% to ~90%.

• Specificity (avoiding false alarms) rose from ~81% to ~89%.

• Overall accuracy (AUC) increased from 0.87 to 0.95.

  
  

In VIOLA-AI detection of intracranial hemorrhage, false positives often arise from motion artifacts or intracranial calcifications, while false negatives occur when subtle or small bleeds in anatomically challenging regions yield probabilities below the 0.5 detection threshold. Addressing these cases is crucial for enhancing models sensitivity, reinforcing the importance of specialist feedback in clinical use.

In other words, the AI became better at spotting subtle hemorrhages while also reducing unnecessary alerts.

Why it matters

Detecting brain hemorrhage is one of the most urgent and critical tasks in neuroradiology. Missing even a small bleed can have serious consequences, while false positives can slow down workflows.

What this project shows is that AI does not have to remain a static “black box.” With tools like VIOLA-AI, it can become a partner that learns from neuroradiologists, improving accuracy and safety step by step.

Key takeaways

• Even with experienced radiologists, small hemorrhages can sometimes be missed, showing the importance of supportive tools.

• Adaptive AI like VIOLA-AI demonstrates that systems can improve when trained directly (daily) with radiologist feedback, leading to safer care.

• Although AI is a powerful tool, right now it is only being trained on very specific tasks. Its usefulness is real, but narrow. Expertise remains essential, and AI should be seen as an assistant, not a replacement.

Complete study:

Liu Q, Nesvold J, Murugesu E, Røvang M, MacIntosh BJ, Bjørnerud A, et al. Examining deployment and refinement of the VIOLA-AI intracranial hemorrhage model using an interactive NeoMedSys platform [preprint]. arXiv [Internet]. 2025 May 14 [cited 2025 Aug 27]; Available from: https://doi.org/10.48550/arXiv.2505.09380