Computational radiology research
Melkorka Labs builds Melkorka, a platform for quantitative radiology. It gives radiologists and clinicians a set of tools to explore medical images, investigate findings, and work alongside AI to ask deeper questions of the data. Ask a question about one patient, or run it across a thousand to measure, compare, and summarize an entire cohort.
Radiology software has traditionally been built around graphical interfaces for human users. Melkorka makes computation, orchestration, and visualization accessible to both people and software agents, while keeping computed results inspectable and reusable.
The platform
Our platform, Melkorka, is built for quantitative radiology: a deterministic study orchestrator, a computation kernel, and a native viewer, designed together to let large language model agents and human users compute, inspect, and reason about structure in medical images.
Orchestration
The study coordinator. It takes a program of analyses across a set of patients, dispatches work to the Kernel, and journals what was admitted. A run records the work that was performed; each answer is identified independently by the request that produced it.
Computation
Typed engines over imaging data: tissue models, geometry, topology, quantitative maps. Each answer is immutable, content-addressed, and memoized.
Inspection
A native DICOM viewer. Multi-planar reformatting, freely oriented oblique slicing, linked viewports, and computed models drawn directly over the source images. It is interactive at full fidelity for user inspection and fully scriptable for agentic control.
Agents
Modern language-model agents can reason across complex workflows and operate software directly. Melkorka exposes the imaging stack accordingly: computation, orchestration, and visualization are available programmatically as well as interactively.
Agents reach Melkorka through MCP and Agent Skills. MCP provides a standard interface for calling Melkorka tools, while Agent Skills package the instructions needed to use them.
Users can work through the large language model of their choice without needing to know Melkorka Loom or the Kernel directly. The agent can run computations, query imaging data, and carry out workflows across a single patient or an entire cohort, then bring the results into the Viewer for review alongside the original images.
Method
Melkorka identifies results the way Git and Nix identify objects: identity comes from content and provenance rather than filenames or mutable locations.
An answer is identified by the request that produced it. Two studies that ask the same question of the same data get the same answer: found, not recomputed.
Our research interest is measurement derived from identified tissue structure rather than arbitrary thresholds. The structure may be established computationally, including with learned models, and then used as the basis for quantitative measurements.
Core assumptions are made explicit. State is fuzz-tested, invariants are checked in the build, and oracles prove a behavior rather than demonstrate it. Agents may reason probabilistically; the computations they invoke produce explicit, inspectable results. These are engineering guarantees about system behavior, distinct from claims of clinical performance.
Stored output can be released when it is no longer needed; the record of how it was produced is kept. Every result knows whether it is stored or dropped and rebuildable, so nothing is ever quietly gone.
Melkorka is the research platform of Melkorka Labs, currently in development. It is not a medical device and is not cleared for clinical use.