midog-mitosis

H&E tumor fields with consensus mitotic-figure annotations, cut from MIDOG++ (503 slides, 7 tumor types, humans and dogs). The tasks config is the benchmark: every row is an image holding at least three mitotic figures (one field, or a pair of fields from one slide), the question asked about it (find every mitotic figure), the expected answer (the coordinates of each) and the rubric the reward follows. The answer is free response and the reward is binary: the listed points match the annotated mitotic figures one-to-one, none missing and none extra, or they do not. It is what the midog-mitosis Verifiers environment serves, as a public benchmark (test split) and an RL environment (train / validation). The warmup config holds the single fields with exactly one figure, in the same format: they are not part of the benchmark, only training material for models too weak to pass any tasks row. fields and candidates are the annotated source tables the tasks are cut from.

Source and license

Everything here derives from MIDOG++ (Aubreville et al., Scientific Data 10, 484, 2023), released by its authors under CC0 1.0 on Figshare (collection 10.6084/m9.figshare.c.6615571). This derivative is released under CC0 1.0 as well. Please cite the original work:

Aubreville, M., Wilm, F., Stathonikos, N., Breininger, K., Donovan, T. A., Jabari, S., Veta, M., Ganz, J., Ammeling, J., van Diest, P. J., Klopfleisch, R., & Bertram, C. A. (2023). A comprehensive multi-domain dataset for mitotic figure detection. Scientific Data, 10, 484. https://doi.org/10.1038/s41597-023-02327-4

Changes made: fields and cell crops were cut from the released slide TIFFs and re-encoded as JPEG (quality 92); cell crops were upsampled 4x (bicubic); annotations were converted from slide to field coordinates; only slides with annotations (503 of the 553 in the released JSON) are used. Labels were not altered.

Configs

tasks (default, the benchmark) and warmup have the same columns: one row per task.

column	meaning
task_id	find:<source_id>
image	the exact image the model sees: a 768 x 768 px field, or two fields of one slide side by side
question	the exact text sent with the image, answer format included
answer	the expected final answer: [[x, y], ...], the center of every mitotic figure, normalized to 0-1000 on each axis
rubric	how the reward of this task is computed
points	the same centers in image pixels
width, height	image size in pixels (768 x 768, or 1536 x 768 for a pair)
n_targets	number of mitotic figures in the image (what the model has to find)
n_imposters	number of annotated non-mitotic look-alikes in the image
parts	the field_ids of the fields rows the image was made from (two for a pair, left then right)
rank	position in the split's fixed pseudo-random order; the environment serves tasks in this order
source_id	parts joined with +
slide, tumor_type, species, scanner, origin	slide metadata

tasks (the benchmark)	with 1 mitotic figure	2	3	4 or more	of which side-by-side pairs	pair halves (figures left+right, sorted)
train: 357	0	0	318	39	0
validation: 64	0	0	57	7	0
test: 512	0	0	452	60	27	1+2+: 27

The answers are the MIDOG++ expert annotations, not model output or heuristics. A task keeps only fields in which every annotated cell is undisputed: the first two experts agreed on every mitotic figure (votes 1,1) and on every look-alike (2,2 or 2,2,2), so the expected answer is not a coin flip that a third expert settled. A task has at least three mitotic figures in tasks and exactly one in warmup, so answering [] is never right. The model is not told how many figures there are, and it passes only by listing every one and nothing else.

Benchmark images are native-resolution fields or same-slide pairs, each containing at least three figures. No benchmark field or figure is reused. A field with exactly one figure is a warmup row, unless it is one half of a benchmark pair: no warmup row shows a field that a tasks row shows.

The 3+ figure criterion is a calibrated difficulty rule, not a preregistered criterion or a performance guarantee. Dense crop and split selection use expert annotations and fixed seeds, not per-task model outcomes. Full builds select 512 test tasks and 64 validation tasks in seeded order, after reserving checksum-pinned seed tasks. Training uses every remaining eligible field on training slides.

The reward is computed by midog_mitosis/scoring.py in the environment from the reply and these columns alone. It requires a one-to-one match to every annotated figure, with none missing or extra, within a 30 px radius. A fixed single-point answer cannot pass a benchmark task. Partial matches are diagnostic metrics, not partial reward.

fields: one row per 768 x 768 px field (192 um at 0.25 um/px, 40x).

column	meaning
field_id	s<slide>_f<coordinate hash>, or a checksum-pinned seed field's id
image	JPEG field
mitoses	[[x, y], ...] field-pixel centers of every consensus mitotic figure in the field
imposters	same for annotated non-mitotic look-alikes (annotated non-exhaustively by the authors)
slide, x0, y0, size	source slide id and the field's top-left corner and edge in slide pixels
kind	how the field was chosen: dense, anchored, mixed or random (below)
tumor_type, species, scanner, origin	slide metadata (lymphoma is the dataset's "lymphosarcoma")

Fields from one slide never overlap. Fixed seed fields are reserved first, then dense fields are packed; up to 2 anchored, 3 mixed and 2 random fields are added in the remaining area:

• dense: at least three mitotic figures, every annotated cell undisputed, selected by deterministic non-overlapping packing.
• anchored: a mitotic figure sits at a uniformly random position inside the field.
• mixed: uniformly random position, kept only if it holds at least one mitotic figure and one imposter and at most 12 annotated cells.
• random: uniformly random tissue.

Objects within 40 px of a border are not listed, and a field is rejected if a mitotic figure would be cut by its border, so the lists match what is fully visible.

candidates: one row per annotated cell, with the same number of mitotic figures and imposters from every slide (up to 8 of each).

column	meaning
candidate_id	s<slide>_a<annotation id>
image	128 x 128 px crop centered on the cell, upsampled to 512 x 512 (JPEG)
label	mitotic figure or imposter (three-expert consensus)
votes	the individual expert labels (1 = mitotic figure, 2 = imposter); disagreements were resolved by a third expert
x, y, slide, tumor_type, species, scanner, origin	as above

Usage

from datasets import load_dataset

tasks = load_dataset("tirandazdylan/midog-mitosis", "tasks", split="test")
row = tasks[0]  # row["image"], row["question"], row["answer"], row["points"], row["rubric"]
fields = load_dataset("tirandazdylan/midog-mitosis", "fields", split="test")
candidates = load_dataset("tirandazdylan/midog-mitosis", "candidates", split="test")

build_report.json in this repository holds the counts, checksums and settings of the build.

Splits

Splits are by slide (every MIDOG++ slide is a distinct case), including all tasks, warmups, fields and candidate crops. This is a custom split, not the official MIDOG++ split: all 111 official test slides stay in test, and additional official training slides are held out until there is capacity for 512 tasks. Remaining slides are assigned to validation until it can supply 64 tasks, then to training. Selection uses seeded, round-robin tumor-type order and annotation-only crop capacity. The split manifest is recorded in build_report.json.

Test covers 173 of 503 slides (34%). An evaluated model's training slides must be disjoint from this release's held-out slides. All configs follow these assignments; using a different training split does not establish held-out validity. Task order is seeded and field ids do not encode a field's content.

config	split	rows	slides
candidates	train	3338	315
candidates	validation	240	15
candidates	test	2286	173
fields	train	2023	315
fields	validation	188	15
fields	test	1531	173
tasks	train	357	49
tasks	validation	64	14
tasks	test	512	106
warmup	train	543	251
warmup	validation	37	14
warmup	test	292	144

Intended use and limitations

Evaluating and training vision-language models on mitotic-figure reading. Not for clinical use.

• The answers are public in this dataset. The midog-mitosis environment runs every attempt in a fresh sandbox that never receives the answer, with outbound traffic blocked except the model route; do the same when you evaluate an agent that can execute code, and don't evaluate a model trained on the test slides.
• Mitotic-figure labels are hard: the paper reports that the first two experts disagreed on about 20% of candidates (19.98%, its Table 3), resolved by a third. Tasks keep only fields without a disputed annotated cell. Imposters were annotated non-exhaustively, and a mitosis the experts missed would count against a model that finds it.
• A pair shows two non-adjacent fields of one slide side by side, as its question states.
• 7 tumor types, 4 scanner models (as named in the official split file), 4 laboratories; not a substitute for external validation.

Built by generators/midog-mitosis/build_dataset.py (seed midog-mitosis-hard-v1). build_report.json records the settings, the Figshare MD5 of every slide, the SHA-256 of the annotation and split files, and the library versions used.