Micro-OD

Micro-OD is a few-shot microscopy object detection benchmark. It aggregates four publicly available biological imaging datasets across distinct microscopy domains and cell types, and packages them into a standardised format designed for evaluating vision models — in particular, large vision-language models (VLMs) — under few-shot, in-context prompting conditions.

Motivation

Microscopy object detection is a challenging setting for general-purpose vision models: images are domain-specific, class vocabularies are narrow but fine-grained, and labelled data is scarce. Micro-OD is designed to probe how well a model can detect cells and parasites in a new domain when given only a handful of annotated example images at inference time — without any fine-tuning.

Dataset Splits

The dataset contains two splits that together constitute the few-shot evaluation protocol:

Split	Role	Images per sub-dataset	Total images
example	Few-shot support set	10	40
test	Evaluation query set	53	212

Evaluation protocol: For each sub-dataset, a model may be provided with up to 10 example images (from the example split) as in-context demonstrations. It is then evaluated on each of the 53 test images in the test split. No fine-tuning on the example images is assumed — they serve solely as few-shot context.

Sub-datasets

Sub-dataset	Domain	Classes	Original size	Format	Source
BBBC	Bright-field blood smear; malaria parasite detection	Red Blood Cells, Trophozoite Cells, Ring Cells, Gametocyte Cells, Schizont Cells, White Blood Cells	1,328 images	PNG	Broad Bioimage Benchmark Collection
BCCD	Peripheral blood smear; blood cell counting	Red Blood Cells, White Blood Cells, Platelets	364 images	JPG	BCCD Dataset
LIVECell	Phase-contrast live cell imaging (RatC6)	Spindle Cells, Polygonal Cells, Round Cells	420 images	PNG	LIVECell (images); annotations in-lab
NIH-3T3	Phase-contrast mouse fibroblast imaging	Polygonal Cells, Spindle Cells, Round Cells	63 images	PNG	In-lab collection and annotation

Folder Structure

Micro-OD/
├── data/                           # Generated Parquet files (HuggingFace viewer)
│   ├── example.parquet             # 40 rows — few-shot support set
│   └── test.parquet                # 212 rows — evaluation query set
│
├── example/                        # Few-shot support set (raw files)
│   ├── BBBC/
│   │   ├── annotation.jsonl        # Bounding-box annotations
│   │   ├── images/                 # 10 PNG images
│   │   └── images_overlay/         # 10 images with bounding boxes drawn
│   ├── BCCD/
│   │   ├── annotation.jsonl
│   │   ├── images/                 # 10 JPG images
│   │   └── images_overlay/
│   ├── LIVECell/
│   │   ├── annotation.jsonl
│   │   ├── images/                 # 10 PNG images
│   │   └── images_overlay/
│   ├── NIH-3T3/
│   │   ├── annotation.jsonl
│   │   ├── images/                 # 10 PNG images
│   │   └── images_overlay/
│   └── stat.txt                    # Split-level statistics
│
└── test/                           # Evaluation query set (raw files)
    ├── BBBC/
    │   ├── annotation.jsonl
    │   ├── images/                 # 53 PNG images
    │   └── images_overlay/
    ├── BCCD/
    │   ├── annotation.jsonl
    │   ├── images/                 # 53 JPG images
    │   └── images_overlay/
    ├── LIVECell/
    │   ├── annotation.jsonl
    │   ├── images/                 # 53 PNG images
    │   └── images_overlay/
    ├── NIH-3T3/
    │   ├── annotation.jsonl
    │   ├── images/                 # 53 PNG images
    │   └── images_overlay/
    └── stat.txt                    # Split-level statistics

Each images_overlay/ folder contains copies of the images with ground-truth bounding boxes rendered on top, useful for visual verification.

Annotation Format

Annotations are stored as JSON Lines (.jsonl) files — one JSON object per line, one line per image.

{
  "image_path": "images/<filename>",
  "bbox": {
    "<class_name>": [
      [[x_min, y_min], [x_max, y_max]],
      [[x_min, y_min], [x_max, y_max]]
    ],
    "<class_name>": [
      [[x_min, y_min], [x_max, y_max]]
    ]
  }
}

Coordinate convention:

• All coordinates are in pixel space.
• Each bounding box is represented as two points: [x_min, y_min] (top-left corner) and [x_max, y_max] (bottom-right corner).
• A class key is present only if at least one instance of that class appears in the image.

Concrete example (from example/BCCD/annotation.jsonl):

{
  "image_path": "images/BCCD_example_1.jpg",
  "bbox": {
    "Red Blood Cells": [
      [[201, 223], [314, 322]],
      [[1, 252], [89, 357]],
      [[203, 336], [292, 441]]
    ],
    "White Blood Cells": [
      [[211, 4], [338, 132]]
    ],
    "Platelets": [
      [[330, 442], [373, 480]]
    ]
  }
}

Usage

To load the dataset:

from datasets import load_dataset

ds = load_dataset("stumbledparams/Micro-OD")

# Access splits
example_split = ds["example"]   # 40 images — few-shot support set
test_split    = ds["test"]       # 212 images — evaluation query set

# Each row contains:
#   image       — PIL image
#   image_id    — "<subdataset>/images/<filename>"
#   subdataset  — one of: BBBC, BCCD, LIVECell, NIH-3T3
#   objects     — dict with keys:
#                   bbox     : list of [x_min, y_min, width, height]  (COCO format, float32)
#                   category : list of int  (ClassLabel index; decode with int2str)

row = test_split[0]

# Image (PIL.Image)
image = row["image"]

# Bounding boxes and category indices
bboxes     = row["objects"]["bbox"]      # list of [x_min, y_min, width, height] (float32)
categories = row["objects"]["category"]  # list of int (ClassLabel index)

# Class names in ClassLabel index order (alphabetically sorted)
CLASS_NAMES = [
    "Gametocyte Cells", "Platelets", "Polygonal Cells", "Red Blood Cells",
    "Ring Cells", "Round Cells", "Schizont Cells", "Spindle Cells",
    "Trophozoite Cells", "White Blood Cells",
]

for bbox, cat_idx in zip(bboxes, categories):
    x_min, y_min, width, height = bbox
    label = CLASS_NAMES[cat_idx]
    print(f"{label}: [{x_min:.1f}, {y_min:.1f}, {width:.1f}, {height:.1f}]")

Note on bbox format: The Parquet files store bboxes in COCO format [x_min, y_min, width, height] as float32. category is stored as a ClassLabel integer index. The raw annotation.jsonl files use [[x_min, y_min], [x_max, y_max]] (top-left / bottom-right pixel coordinates) — see Annotation Format.

Dataset Statistics

Detailed per-class statistics are available in example/stat.txt and test/stat.txt. Summaries are provided below.

Test Split — 212 images

Sub-dataset	Images	Classes	Total boxes	Boxes/image (mean)	Boxes/image (range)
BBBC	53	6	4,000	75.5	19–135
BCCD	53	3	952	18.0	9–30
LIVECell	53	3	223	4.2	1–15
NIH-3T3	53	3	376	7.1	1–14
Total	212	10	5,551	—	—

Example Split — 40 images

The Support-Spread Score (SS) is a composite metric reflecting both class coverage (fraction of classes represented in the sample) and class balance (how evenly instances are distributed across represented classes). Higher is better; a score of 1.0 indicates perfect coverage and balance.

Sub-dataset	Images	Total boxes	Boxes/image (mean)	Support-Spread Score
BBBC	10	734	73.4	0.136
BCCD	10	78	7.8	0.680
LIVECell	10	40	4.0	0.763
NIH-3T3	10	62	6.2	0.612
Total	40	914	—	—

The low SS for BBBC (0.136) reflects the extreme dominance of Red Blood Cells in the malaria dataset, which makes it difficult to achieve a balanced 10-image sample across all 6 classes.

Class Inventory

Class	Sub-dataset(s)	Test boxes
Gametocyte Cells	BBBC	24
Platelets	BCCD	159
Polygonal Cells	LIVECell, NIH-3T3	417
Red Blood Cells	BBBC, BCCD	4,427
Ring Cells	BBBC	34
Round Cells	LIVECell, NIH-3T3	24
Schizont Cells	BBBC	10
Spindle Cells	LIVECell, NIH-3T3	158
Trophozoite Cells	BBBC	193
White Blood Cells	BBBC, BCCD	105

Note that Polygonal Cells, Round Cells, and Spindle Cells appear in both LIVECell and NIH-3T3 but describe morphologically similar — not biologically identical — phenotypes in different cell lines.

Attribution

Micro-OD combines images and annotations from multiple sources. Please credit the original sources as appropriate:

• BBBC (malaria): Ljosa, V., Sokolnicki, K. L., & Carpenter, A. E. (2012). Annotated high-throughput microscopy image sets for validation. Nature Methods, 9(7), 637. https://bbbc.broadinstitute.org/

• BCCD: Shenggan. BCCD Dataset. GitHub. https://github.com/Shenggan/BCCD_Dataset

• LIVECell (images): Edlund, C., et al. (2021). LIVECell — A large-scale dataset for label-free live cell segmentation. Nature Methods, 18(9), 1038–1045. https://doi.org/10.1038/s41592-021-01249-6. The morphology-based bounding-box annotations used in Micro-OD were produced in-lab and are not part of the original LIVECell release.

• NIH-3T3: Images and bounding-box annotations are an in-lab collection and are not sourced from a public dataset.