Medical AI · Medical imaging
We develop and evaluate segmentation, detection and classification models for medical images, including spectral imaging. The work covers the whole chain: data from several sites and devices, labels that experts disagree on, metrics that reflect clinical errors, and images that must be de-identified before they move.
Segmentation, detection & classification
The output follows the clinical question: an outline for volumes or planning, a box or point for finding lesions, a label for triage.
Imaging is part of our Medical AI practice. For how imaging models reach production, see computer vision in production.
Pixel- or voxel-level outlines of organs, lesions or structures, in 2D or 3D. Self-configuring U-Net baselines such as nnU-Net are strong starting points, and give any custom architecture a fair benchmark to beat.
nnU-Net, Nature Methods →Finding and localising lesions or findings, often small and rare. Evaluation counts lesions, not pixels: how many are found, how many false alarms per scan, and at what size.
Image- or study-level labels for screening, triage or prioritisation. Thresholds are set with clinicians, weighing the cost of a missed case against the cost of a false alarm.
Medical datasets are often small. Transfer learning, self-supervised pretraining on unlabelled scans and careful augmentation can help. Only a test set from patients and sites the model never saw shows whether they did.
Multi-site & scanner shift
Vendor, protocol, reconstruction, resolution and patient mix all change the pixels. A model trained at one site often loses accuracy at the next.
Models can pick up markers of the hospital or device instead of the disease. In one chest X-ray study, pneumonia models did worse on another hospital's images in three of five comparisons, and a network could tell which hospital system an image came from.
Zech et al., PLOS Medicine 2018 →Always split by patient. Hold out whole sites or scanners when the question is transfer. Report performance per site and per device, not only pooled.
Intensity normalisation, resampling to a common spacing, and augmentation that mimics real acquisition differences. These narrow the gap. They do not replace testing on the target site.
Scanner software updates and protocol changes shift inputs silently. We define input checks and image statistics to monitor, so drift is seen before performance falls.
Clinical AI in production →Labels & ground truth
Two readers rarely draw the same outline. Their disagreement can be as large as the error you are trying to measure.
Where the budget allows, several annotators label the same cases. Agreement between readers sets a realistic ceiling: a model that matches readers as well as they match each other is near the limit of the labels.
Majority voting, consensus review or STAPLE-style estimation combine several outlines into one reference, and estimate how reliable each reader is.
STAPLE, IEEE TMI 2004 →A written protocol with examples, edge cases and a pilot round reduces disagreement before it reaches the data. We review the hardest cases with your clinicians.
Where possible, labels come from a stronger source than the image itself: pathology, follow-up or outcome. Labels read from the image alone are reported as such.
What we measure
Overlap scores are easy to compute and easy to misread. We choose metrics that match the clinical question, following published recommendations such as Metrics Reloaded.
Dice and IoU summarise overlap. They punish errors on small structures far more than on large ones, and ignore where the error is. A high average Dice can hide a missed small lesion.
Hausdorff distance, often its 95th percentile, and average surface distance show how far an outline strays. That matters when the output is used for planning or measurement.
Detection sensitivity, false positives per scan and results by lesion size. A model that finds every large lesion and misses the small ones may fail the clinical need.
Volumes, diameters or counts derived from the output, compared with what clinicians measure. Reader studies where the question is whether the model helps a reader.
Metric choice follows Metrics Reloaded and its companion paper on metric pitfalls. Results are reported per site and subgroup, with confidence intervals. We do not quote target scores before seeing the data.
De-identifying images
DICOM files hold names, IDs, dates and device details in their tags. Some identifiers sit in the pixels themselves.
The DICOM standard defines confidentiality profiles: which attributes to remove, replace or keep, with options to retain dates, device details or longitudinal links when a study needs them.
DICOM PS3.15, Annex E →Ultrasound, screen captures and scanned documents can have names or IDs written into the image. Header cleaning does not touch them. We detect text in the pixels, mask it, and check a sample by eye.
A 3D head scan can be rendered into a face. In one study, face-recognition software matched MRI-based reconstructions to participants' photos. Defacing tools remove facial features, at some cost to later analysis.
Schwarz et al., NEJM 2019 →Vendor private tags, study descriptions and embedded reports can contain identifiers. They need review per vendor and modality, not one generic script.
Clinical data de-identification →How an engagement runs
An imaging project is decided early: by the reference standard, the annotation protocol and how the data is split. Our process reflects that.
Agree the clinical question, the output (outline, box or label), the reference standard and the metric that reflects clinical success.
De-identify, check image quality and acquisition metadata, write the annotation protocol, and split by patient and site before any training.
Start from a strong baseline, then test each change against it on held-out patients. Review failures by site, scanner and lesion size.
Test on held-out sites, report per-site and subgroup results in a report structured around CLAIM, and hand over code your team can retrain.
We produce development and validation evidence. We are not regulatory consultants, and engaging MLAIA does not imply any regulatory status for your product.
Sources
Before we start
Not on its own. Dice favours large structures and ignores where errors fall. We pair it with boundary distances, lesion-level detection and the measurements clinicians actually use.
Usually acquisition shift: different scanners, protocols or patient mix, sometimes made worse by shortcuts the model learned at the first site. We test per site, find what changed, and choose between harmonisation, more varied training data and local fine-tuning.
It depends on the task, the variability and how small the target is. We look at a pilot set and at agreement between readers before estimating. Fewer cases labelled well usually beat many labelled loosely.
No. Identifiers can be burned into the pixels, hidden in private tags or free-text fields, or reconstructed, as with faces in head scans. We clean headers, check pixels and review a sample by eye.
Yes. We can work inside your environment under your access controls, or on de-identified exports prepared for the task. This is agreed with the data owner before work starts.
No. We produce the development and validation evidence, structured so it can support regulatory work. We are not regulatory consultants, and engaging MLAIA does not imply any regulatory status for your product.
Medical AI · Medical Imaging AI
Tell us the clinical question, the images and labels you have and what a reviewer would need to see. We’ll tell you whether it’s feasible and what the first step would be.
Please don’t include confidential datasets, credentials or patient information.