Skip to content

Stage 0 - Chest X-ray dataset loaders and case table (MIMIC-CXR primary, CheXpert, NIH) #4

Description

@sebasmos

Goal. Load the three imaging substrates into one shared case schema, keyed so pairing can swap images within a patient.

Scope / what to build.

  • Loaders for MIMIC-CXR (primary; radiographs plus paired free-text reports), CheXpert (its device labels power a later natural-cue validation arm), and NIH ChestX-ray14 (open replication arm). Reuse a standard chest X-ray data library that already exposes all three with aligned labels rather than parsing each raw layout by hand.
  • A unified case table: one row per image with a stable case id, patient id, image handle, report text (where available), and label. Keyed so the pairing step can pick a same-patient swap image.
  • Record the dataset revision in the manifest.

Reuse vs bespoke. Reuse the data library's loaders and label alignment. Bespoke: the case-table join and the same-patient key.

Definition of done. Each dataset loads into the shared schema; a smoke test builds the case table for a small sample from each; same-patient grouping is exposed to the pairing step.

Model backend. For now every agent uses the Gemini API (multimodal). All model access goes through one gateway wrapper, so the roster can be extended to other model APIs later without changing any experiment code. Gemini is the default for now; no fine-tuning, models are used off-the-shelf.

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions