XDMD

Code for imaging. AI models that keep working in the clinic.

What you see here is a CT scan of a patient with a suspected stroke. We build the AI model that analyzes a scan like this. And not only CT. We work on any medical image, whether it is an MRI, an ultrasound or a pathology slide.

We build the software that makes AI work in clinical practice, the model and everything around it. That starts with the data, and with understanding what we see in it. Collecting representative data and good annotations is the hardest part, as a model is only as good as the examples it learns from.

An AI model is never finished. When the scanners, the clinical insights or the patient populations change, the data drifts away from what the model has learned, and its performance degrades. So a model has to be monitored and updated, and sometimes rebuilt. AI development in this domain never stops.

How we build

Superstack

Our AI development environment

DaVinci

Our cloud-based annotation platform

In production

An H&E section with the pathologist's marker rings, above three unstained dissection slides of the same tissue block. The marked region, carried onto each slide by the registration, is outlined in amber.
The H&E (top) with the region the pathologist marked, in amber, and three unstained dissection slides of the same block, with tissue in green and paraffin in sand. For Xyall we built a pipeline with Superstack, trained on data annotated in DaVinci. It segments each slide, registers the H&E onto each unstained slide and carries the marked region over: this is where their tissue dissection robot collects the tumor tissue. The pipeline runs inside the robot, on its embedded hardware, in their clinical workflow. On a representative test set, 99% of its segmentations and 93% of its registrations passed expert review; the next release aims at 99.9% in the field. 5 mm
Rashindra’s sharp mind makes working with him a delight. Rather than relying on off-the-shelf tools, he challenged us to truly understand our data. XDMD’s approach led to a custom solution that significantly improved our image registration process.
Dr. Sven Lafebre
Senior Software Architect at Xyall, The Netherlands

Trusted by

Xyall

Slide registration for a tissue-collecting robot

Canon

Vessel segmentation and collateral scoring in CTA

Radboudumc

The first brain tissue segmentation in 4D CTA

How we work

Understand the problem first

We find out what the problem really is, and why things are as they are. Then we zoom out and ask whether it is the right problem to solve, before we write a single line of code.

The best code is no code

What we do not build cannot fail and needs no maintenance. What we build is only what the job needs.

Never fail silently

When a case is outside what our software can handle, it says so and stops, instead of quietly giving a wrong result.

Watch every model in use

Scanners, protocols and patients change. We check each new case against what the model learned from, and update the model before its results degrade.

Who we are

Our team and advisors have spent decades in academic hospitals and in industry, and have written more than 1000 peer-reviewed publications. We built this expertise before and during the rise of AI, and today we use AI to build AI.
Meet the team

Rashindra is one of the most knowledgeable engineers in medical imaging and machine learning. If you look for solid solutions for your image analysis problems, look no further.
Prof.dr. Bram van Ginneken
Professor of Medical Image Analysis at Radboudumc, Founder at Plain Medical, Founder at Thirona, The Netherlands

Work with us

Have a problem in medical imaging? Tell us in a few lines what the problem is and what data you have. Within one week we tell you how we would approach it, and whether we are the right team to solve it.

Get in touch