Building a Medical AI Startup in Turkey
Building an orthopaedic AI startup in Istanbul — architecture decisions, regulatory navigation, and Turkey's medical AI ecosystem.
Why Turkey, Why Now
Turkey's position in the medical AI landscape is unique. A population of 85 million generates substantial clinical volume — Turkish hospitals perform over 70,000 total knee arthroplasties annually. The regulatory environment, while rigorous, is actively adapting to AI through TİTCK (the Turkish Medicines and Medical Devices Agency), which has aligned with EU MDR and is developing specific guidance for AI-as-medical-device. And the cost structure for a pre-revenue startup — engineering talent, cloud infrastructure, office space — is significantly more favourable than Western Europe or the US.
Salnus was founded in Istanbul in 2025 with a specific thesis: orthopaedic surgeons need AI tools that integrate into their existing workflow, not standalone products that create additional steps. Our first year has validated this thesis and taught us several lessons worth sharing.
Technical Architecture Decisions That Defined Us
Two architectural choices made early have shaped everything that followed.
Client-side AI processing was a deliberate decision to prioritise data privacy by architecture rather than compliance by process. By running AI inference in the surgeon's browser using ONNX Runtime Web, we eliminated GPU server costs, simplified our KVKK obligations, and gave surgeons confidence that their patient data never leaves their device. This decision has shaped our technical architecture — no per-inference cloud compute charges. Next.js on Vercel keeps hosting near zero. The result: a production-grade medical AI platform that costs less than $50/month to operate.
DICOM-native design meant building on Cornerstone3D from day one, rather than converting DICOM to JPEG and losing clinical metadata. This was more complex initially but gave us calibrated measurements, proper windowing, and MPR support — features that distinguish a clinical tool from a demo.
Working with Surgeons
The single most important lesson from our first year: surgeons are not users — they are collaborators. The difference is not semantic.
A user receives a finished product and provides feedback. A collaborator shapes the product from the earliest design stage. Our DICOM viewer, AI pipeline, and clinical reporting system were built in direct dialogue with practising orthopaedic surgeons who defined the requirements based on their daily clinical workflow.
This collaboration model extends to our business structure. For custom clinical software projects, we operate on a shared intellectual property and revenue-sharing basis with surgeon partners. For our 3D-printed PSI service, the surgeon is involved in every planning decision. This is not altruism — it is the only way to build products that actually get used in the operating room.
Navigating KVKK
Turkey's KVKK (Kişisel Verilerin Korunması Kanunu) is substantively aligned with GDPR, with some procedural differences. For a medical AI company, the key obligations are: explicit consent for health data processing, appointment of a data controller, registration with the VERBİS registry, and implementation of appropriate technical measures.
By choosing client-side processing, we eliminated the most complex KVKK obligations around server-side health data storage. By pseudonymising patient IDs with SHA-256 hashing, we minimised the personally identifiable information in our system. By designing data processing agreements that clearly delineate what Salnus does and does not receive, we made the legal relationship with partner institutions straightforward.
The lesson: in regulated industries, compliance is not a checkbox to complete after building the product. It is a design constraint that should inform every architectural decision from the beginning.
The Research Foundation
Academic credibility matters in medical AI — perhaps more than in any other software domain. Surgeons evaluating an AI tool ask two questions: "where was this published?" and "who validated it?"
Our OJSM publication on 3D-printed patient-specific guides for knee reconstruction established Salnus as a research-active organisation, not just a software company. Our systematic AI experiments — 21 experiments across 11 architectures, with honest reporting of both successes and limitations — demonstrate scientific rigour.
The lesson: publish early, publish honestly. A paper showing 70% accuracy with transparent methodology builds more clinical trust than marketing materials claiming 99% accuracy on an undisclosed dataset.
What Comes Next
Our roadmap for 2026 focuses on three parallel tracks: expanding our AI model accuracy through larger training datasets, deploying the platform for pilot use at partner clinics, and building the regulatory documentation needed for eventual CE marking.
We are actively seeking orthopaedic surgeon partners for clinical validation pilots, hospitals interested in evaluating AI-assisted DICOM viewing, and academic collaborators for joint research projects. If any of these align with your interests, contact our team.
Salnus Medikal Yazılım ve Cihaz Teknolojileri San. Tic. A.Ş. is based in Istanbul, Turkey.
Reviewed by the Salnus biomedical engineering team.