External Tools Evaluated

Third-party software read at source level and weighed against what the group already runs — what each one is for, and the single idea worth taking from it

← Home

What this is

A register of other people’s repositories that have been read here — at source level rather than from their READMEs — and recorded so that nobody runs the same evaluation twice. Each has a page in the research wiki carrying the full reading; this page is the index to them.

The question asked of every one of these was not what does this tool do but what does it do that the tools already in the stack do not. That framing is why most of them are not in routine use, and also why almost all of them left something behind: an idea, a protocol decision, or a number worth reusing, even where the software itself was the wrong fit. Those are what this page lists.

Registration and slide handling

  • VALIS — registers a whole series of slides from one block rather than a pair, and exposes a public function for warping point coordinates. The use case here is drawing a tumour boundary once on the H&E and carrying it onto the IHC instead of redrawing it, which is what QuPath Annotation Workflow has wanted and never had.
  • ScanTools — label and macro extraction, barcode decoding, pyramid hashing and scan dates. Its standing value is as a read-only crawler over an image share when the question is what have we actually got, which is the prerequisite behind Slide Metadata Extraction.
  • WSITools — patch extraction, tissue detection and annotation parsing. Read for its registration stage, which turned out to be translation-only; the durable output was the pair-versus-series distinction now written up on Cross-Stain Registration.

Survival and prognostic modelling

  • MultiSurv — a multimodal survival model combining images, molecular data and clinical variables. The transferable part is its discrete-time survival loss, which handles the missing modalities that make Multimodal Fusion hard in practice.
  • SurvivalNet — a deep Cox package from an earlier era. Two things survive it: its per-patient gradient attribution, which is the worked example behind Gradient-Based Feature Attribution, and its demo cohort, which is this repo’s clearest demonstration of how Concordance Index inflates when a model can tell which disease the patient has.

Spatial proteomics

Neither is runnable here — there is no multiplex imaging in the department — and both are worth reading for one design decision each.

  • KRONOS — a foundation model with a marker embedding, so a channel is identified by which marker it is rather than by its position in the stack. That single change is the whole of “panel-agnostic”, and it is the clearest illustration of the argument on Spatial Proteomics.
  • CORAL — a full pipeline whose transferable habits are six encoders behind one flag on identical preprocessing (which is Encoder Benchmarking done properly), and writing tissue masks as GeoJSON specifically so they can be hand-corrected in QuPath — an independent arrival at this group’s own interchange decision.

Reading behaviour and agents

  • Pathology-CoT — companion code for an agent paper that navigates a slide the way a pathologist does. Its behaviour-discretisation thresholds are worth reimplementing, and it is the concrete example behind Agentic Slide Navigation: the scarce training signal moves from labelled tissue to recorded expert behaviour.
  • MouseTracks — a desktop input-heatmap tool. What it produced here was the durable part: the two-property test on Reading-Behaviour Capture — positions in image space with the zoom factor, and preserved visit order — which decides whether any recorded viewing data can answer a reading-behaviour question at all.

LLM tooling

  • askLLM — an LLM question-answering interface inside jamovi, and the best-measured example here of grounding suppressing hallucination: sending the real catalogue instead of letting the model infer one took it from three structural fabrications to eighteen of eighteen correct. That trick, and the mechanical check it enables, is what LLM Report-Quality Checking should try next.
  • OpenMed — a healthcare-AI package whose contribution was a pattern rather than a dependency: the release gate, now written up as De-identification Release Gating — an assertion that must pass, emitting a tamper-evident artifact, rather than a step that merely runs.

Platforms and catalogues

  • MIKAIA — a commercial platform worth knowing about for exactly two capabilities the QuPath stack does not have: ISH quantification and cell-neighbourhood analysis. It has a REST API, so it could supplement rather than replace.
  • Awesome-AI4DigitalPathology — a curated index of 453 computational-pathology works whose fifteen sections map almost onto this wiki’s own structure, which makes it usable as a gap finder rather than only a reading list. Two whole areas missing here were found that way and are now written: Federated Learning in Pathology and Cervical Cytology Screening.

Front ends, and one out-of-domain tool

  • Valis-GUI — a desktop wrapper for the registration package above. Read together with it, the useful conclusion is about shape rather than about this particular application: a QuPath menu entry that shells out to a Python process, the pattern already working in QuPath Hepatocyte Extension, is a cheaper way to give a pathologist a button than maintaining somebody else’s desktop application.
  • OncoLens — recorded because its name and description do not say what it is. It fits a growth–death ODE to serial RECIST lesion measurements and plots them against a treatment timeline, which is radiology data and mathematical oncology: no slide, no image, no report text. It is worth one look for the contrast it draws with everything else here — a mechanistic model of one patient’s own trajectory, where all of this group’s time-to-event work is statistical across a cohort. See Survival Analysis Workflow.

How a tool gets onto this page

  1. Read the source, not the README — argument parsers, entry points, defaults.
  2. Ask what it does that QuPath Tooling or the jamovi stack does not already cover.
  3. Write the verdict back into the hub page, so the next evaluation can answer step 2.
  4. Record the one idea worth keeping, even when the answer is no.