Skip to content

README.md

Latest commit

 

History

History
107 lines (85 loc) · 4.35 KB

README.md

File metadata and controls

107 lines (85 loc) · 4.35 KB
Raw

PBSInstr and PBSBench annotations

This directory contains publication-ready QA annotations. Source images are not relicensed here.

data/
├── PBSInstr/
│   ├── cell_captions.jsonl
│   ├── cell_train.jsonl
│   ├── cell_val.jsonl
│   ├── slide_train.jsonl
│   └── slide_val.jsonl
└── PBSBench/
    ├── cell_id_test.jsonl
    ├── cell_ood_test.jsonl
    └── slide_id_test.jsonl

Released annotation counts

File QA pairs
PBSInstr/cell_train.jsonl 20,985
PBSInstr/cell_val.jsonl 3,240
PBSInstr/slide_train.jsonl 1,142
PBSInstr/slide_val.jsonl 144
PBSBench/cell_id_test.jsonl 3,289
PBSBench/cell_ood_test.jsonl 4,495
PBSBench/slide_id_test.jsonl 138

PBSInstr/cell_captions.jsonl contains 28,970 Phase-1 cell-caption records. The same manifest carries each cell's containing patch_path and is therefore also the Phase-2 cell-patch pair manifest.

The released QA counts differ slightly from those reported in the paper because several annotation errors were corrected during final data preparation, which shifted the distribution across the published splits.

Each JSON line contains stable id, human-readable image_id, relative image_path, dataset-native source, level, split, domain, task, question_type, question, answer, and optional options. Question types are true_false, multiple_choice, fill_blank, and open.

Split granularity

The cell ID benchmark is a cell-crop-level holdout. Its cell images are disjoint from PBSInstr cell images, but their parent S-BIAD440 WSIs are not: all 189 source slides represented in the cell ID test set also contribute different cell crops to cell training. It must not be described as a patient/slide-level holdout.

The slide ID benchmark is a WSI-level holdout defined by slide_split.csv: PBSInstr contains 192 training and 24 validation slides, while 23 different slides form PBSBench. No WSI appears in more than one of these splits.

Linking QA records to images

id identifies a QA pair; image_id identifies the image shared by one or more QA pairs. For an in-domain cell, the identifier has the form:

S-BIAD440:<slide>:<patch-x>_<patch-y>:<Cellpose-object-index>

Its source object records the S-BIAD440 WSI path, level-0 patch coordinates, patch size, object index, and curated bounding box. pbsbench-prepare-data uses this locator to generate the record's image_path exactly. Slide records use S-BIAD440:<slide> and point directly to the corresponding downloaded WSI.

OOD records use provider-native identifiers rather than local download wrappers: TCIA records use paths such as NGS/NGS_5602.tiff, Kaggle APL records use paths below All/All/, and WBC LISC records use the unique source filename. The materializer searches below the user-supplied extraction root and recognizes filenames rewritten by Roboflow exports.

Example:

{
  "image_id": "S-BIAD440:III_1:18432_95744:0016",
  "image_path": "cells/S-BIAD440/III_1/18432_95744_0016.png",
  "source": {
    "dataset": "S-BIAD440",
    "path": "images/III_1.tiff",
    "patch_x": 18432,
    "patch_y": 95744,
    "patch_size": 512,
    "object_index": 16,
    "bbox": {"left": 382, "top": 30, "right": 442, "bottom": 99}
  }
}

The released PBSInstr and PBSBench QA manifests are versioned directly in this repository under data/PBSInstr/ and data/PBSBench/. Source images are acquired separately and are not redistributed here:

See docs/PREPROCESSING.md for path-based materialization commands and THIRD_PARTY.md for upstream licenses and notices.

License

The curated PBSInstr and PBSBench question-answer annotations and associated metadata are available under CC BY 4.0. This license does not cover source microscopy images or other third-party assets.