Computer Vision · Healthcare
AI in Medical Imaging: Where It Helps Hospitals and MedTech
MRI and CT scans hold more information than a team can evaluate in the time available. AI takes over time-consuming steps: outlining structures, marking suspicious areas, checking image quality. The assessment stays with doctors. Two projects show what that looks like in practice.
Julia Rose
Key takeaways
- AI in medical imaging supports segmentation, flagging findings that are hard to see, quality control and workflows.
- At apoQlar, segmenting anatomical structures for surgical planning takes seconds instead of hours.
- For the Hamburg Epilepsy Center, a model detecting focal cortical dysplasias reached 90% sensitivity and 70% specificity.
- It takes well-annotated data, traceable pipelines and operation in your own infrastructure where patient data requires it.
Why AI helps in medical imaging
Medical imaging is central to diagnosis, monitoring and treatment planning. The number of scans keeps growing, and many steps around them take time: outlining structures slice by slice, looking for subtle changes, comparing scans. Deep learning models can take over or prepare these steps. They work consistently, do not tire and deliver results in seconds.
The assessment stays with the experts. AI marks, measures and prepares, and the radiologist, neurologist or surgical team makes the decision.
Where AI is used in medical imaging
| Application | What AI takes over |
|---|---|
| Segmentation | outlining bones, vessels and organs in MRI and CT automatically, for example for surgical planning |
| Flagged findings | marking abnormalities that are hard to see, which experts then assess |
| Quality control | checking whether scans are good enough for evaluation before time is lost |
| Workflows | pre-sorting and registering scans and preparing measurements |
| 3D visualisation | building models from segmented structures for planning and patient education |
Two projects in practice
In practice: apoQlar
The VSI HoloMedicine® platform shows surgeons a patient’s anatomy in 3D through an XR headset. Until then, radiologists and technicians traced the structures by hand, slice by slice, for hours per patient. We developed U-Net models, one variant per target structure, for MRI and CT. Segmentation now takes seconds instead of hours, with consistent quality and reproducible pipelines on Azure ML.
In practice: Hamburg Epilepsy Center
Focal cortical dysplasias are a common cause of drug-resistant epilepsy and are often barely visible on MRI. Together with the Epilepsy Center at Evangelisches Krankenhaus Alsterdorf we developed a neural network that marks these lesions. In the prospective study it reached 90% sensitivity and 70% specificity. The results were published in two peer-reviewed papers.
How the model was developed further with continual learning is covered in our article Building AI that learns and adapts.
What these projects need
- Annotated data: scans in which experts have carefully marked the target structures or findings. Their quality determines the quality of the model.
- Clinical collaboration: radiology, neurology or surgery define what should be found and when a result is useful.
- Traceability: datasets, training and evaluation are fully documented, so it can be checked later how a model reached its result.
- Data protection in operation: where patient data requires it, models run in your own infrastructure and documents are anonymised first.
- Fit into the workflow: results have to arrive where people work, for example directly in the planning software.
How hospitals and MedTech companies get started
- Name the bottleneck: which imaging step costs the most time today or leads to inconsistent results?
- Check the data: which scans and annotations exist, and in what quality?
- Prototype on your own data: one model for one structure or finding, measured against expert assessments.
- Build it into the workflow: only once the results convince is the step integrated into existing software.
More applications, from anonymising doctors’ letters to security questionnaires, are on our AI for healthcare and MedTech page.
Relieving imaging work in your hospital or product?
We look at which step suits AI, which data is available and what a first prototype looks like.
Request a process analysisFrequently asked questions
For example automatic segmentation of structures in MRI and CT, flagging findings that are hard to see, quality control of scans, preparing workflows and 3D models for surgical planning.
No. AI marks, measures and prepares. The assessment and the decision stay with the experts.
Scans in which experts have carefully marked the structures or findings of interest. The amount and quality of these annotations determine how good the model becomes.
Yes. Where patient data requires it, the models run on your own servers or in the cloud of the hospital or company.
Julia has been part of theBlue.ai since 2019 and has accompanied the development of AI applications in the enterprise environment since the company’s early days. In her role as Marketing Lead, she works closely with the engineering and consulting teams and makes complex technical topics understandable and accessible for decision-makers.
In her articles, she writes about practical experience from enterprise AI projects, as well as the challenges and opportunities of using AI in companies.