Humboldt-Universität zu Berlin - Mathematisch-Naturwissenschaftliche Fakultät - Institut für Informatik

Der Talk findet auch digital per Zoom statt. Eine Zoom-Einladung finden Sie hier. (nur mit Informatik-Account)

Scientific image analysis increasingly depends on collaborative software routines that must be iteratively developed, tested, shared, and scaled, usually across contributors with different technical expertise. This is particularly difficult for interactive routines, where graphical exploration, heterogeneous dependencies, reproducible execution, and later workflow integration must coexist. Existing infrastructure addresses parts of this problem: package managers distribute
libraries, containers isolate environments, workflow engines scale analyses, and registries improve discovery. However, these approaches do not provide a lightweight executable unit that remains suitable for interactive development while also supporting reproducibility, modularity, and long-term distribution. I address this problem by introducing an executableroutine abstraction, implemented in the open-source framework Album. This abstraction represents a scientific routine
as a self-describing execution unit that combines metadata, typed arguments, environment specifications, and lifecycle hooks for installation, testing, and execution. A complementary catalog abstraction provides decentralized, git-native indexing, sharing, and versioning of such routines. Together, these abstractions separate the scientific routine from any single package, container, or workflow system while preserving the information required to execute, inspect, share, and update it. The central technical contribution is a two-context execution model. A host controller evaluates and validates
the routine manifest, prepares an isolated per-routine environment, and delegates execution to a pinned runner inside that environment. The runner then executes lifecycle hooks under the interpreter and dependencies declared by the routine itself. This separation between control context and execution context enables reproducible post-environment setup, interactive execution, and the composition of routines with incompatible dependencies. It also shifts modularity from
the level of libraries or static workflow steps to independently executable, environment-carrying routine components. I
evaluate this abstraction through two collaborative image-analysis software projects. The first evaluation applies the principle to exploratory fluorescence microscopy analysis in Polarity-JaM. Beyond serving as an application case, this work contributes a target-oriented feature model that organizes measurements across cells, nuclei, organelles, markers, and junction interfaces; circular-statistical readouts including the signed polarity index for cue-dependent polarity; and interactive analysis interfaces through a web application and Napari plug-in. Within this setting, the two-context execution model is used to integrate interchangeable segmentation backends with distinct dependencies behind a common semantic interface. The second evaluation applies the same principle in ARI3D, a software system for interactive quantification of regions in three-dimensional X-ray CT images. Here, segmentation, property extraction, and data viewing are implemented as independently updateable execution units that can evolve outside the core application while preserving interoperability through stable boundary contracts. This evaluates environment-decoupled workflow modularity: individual workflow components can be updated, reinstalled, or replaced without rebuilding the complete software stack, even when their dependency requirements differ. Taken together, the core contribution of this thesis is the design, implementation, and evaluation of a reproducible modular execution layer for collaborative scientific software development, instantiated in Album and evaluated through Polarity-JaM and ARI3D. By turning scientific routines into executable, version-controlled, and cataloged units, this layer bridges the gap between ad hoc research scripts, package managers, containers, and workflow systems. The same abstraction also provides a foundation for LLM-assisted composition of scientific workflows from cataloged executable routines. By exposing the framework’s catalog and execution operations as typed tools, an LLM-enabled agent can discover, inspect, install, parameterize, and execute available routines instead
of generating unconstrained shell commands. This does not replace explicit workflow definitions, but enables a complementary mode of runtime workflow composition in which pipelines are assembled from versioned and reproducible scientific routines.