AIR-LEISH

Overview

AIR-LEISH contains 180 Giemsa-stained microscope images of macrophages infected with Leishmania, with expert-drawn outlines for nuclei, whole cells and the small parasites inside them. It can support cell segmentation, parasite counting and studies of infection.

There are 90 images of THP-1-derived macrophages infected with L. major (Set1), and 90 images of primary human monocyte-derived macrophages infected with L. infantum (Set2). These are laboratory cell preparations, rather than tissue sections.

License: CC BY 4.0. You may use, change and share the data, including commercially, if you credit the creators, keep the license notices and describe your changes.

Files and splits

Split	File	Images
train	train.parquet	180

All images are kept together. The paper describes random training, validation and test sets, but the released files do not identify those assignments or which images share a slide or donor. There is no independent validation or test split in this version. Some images show the same cells in shifted views, confirmed in both infection models. A random image split can therefore put related views into both training and evaluation.

Set1.coco.json and Set2.coco.json preserve the original outlines and source metadata. License and annotation/split notes are included alongside the Parquet file.

Fields

Each row contains one image and all its annotations.

Field	Contents
id	Unique image name, such as air_leish_set1_000
image	Colour image stored as PNG bytes; HWC RGB uint8, values 0–255
inst	Numbered nucleus mask, stored as uint16 PNG bytes
inst_cell	Numbered whole-cell mask, stored as uint16 PNG bytes
inst_parasite	Numbered parasite mask, stored as uint16 PNG bytes
type	Nucleus category mask: 3 inside a nucleus, 0 elsewhere; uint8 PNG bytes
pairs	Suggested nucleus-to-cell matches; several nuclei may match the same cell
objects	Every individual object mask, including overlapping pixels; details below
source_mask	Original combined annotation PNG bytes; null for the one image without this file
source_set, infection_model	Original collection and its cell/parasite model
file_name, source_file_name	Actual archive filename and original COCO filename
source_image_id	Original image number within source_set

Images are 2,709 pixels high × 1,844 wide; all masks use the same HW shape. Zero is background in the numbered masks. Instance number = original annotation number + 1, because the source uses annotation number 0. A nucleus and its cell can have different numbers.

objects is a list with instance_id, source_annotation_id, category_id and rle_counts. The last field holds a complete binary mask in compressed COCO run-length encoding, using the image's height and width. The loading example reads one such mask. Object numbers are scoped to an image and source set.

Labels

Structure	Original objects	Visible in numbered masks
Nuclei	1,634	1,633
Whole cells	1,427	1,427
Parasites	7,994	7,994

Every original object is kept in objects, even when another object overlaps it. The three numbered masks give each pixel one object number within its layer. Where objects in the same layer overlap, the smaller object takes the shared pixels; the larger object number wins equal-area ties. One nucleus is completely covered by another in the numbered mask; both full masks remain in objects. Use objects when you need the full overlapping masks. The original polygon coordinates are also kept in the COCO files.

The source categories are 1 = AM (amastigote parasite), 2 = HC (host cell), and 3 = NU (nucleus). Category 0 is an unused parent category named AM. These are structure categories, not different nucleus cell types. type keeps category 3 for nuclei; it does not combine cells and parasites into one flat mask.

Whole-cell masks include their nuclei. Some cells have several marked nuclei, and some nuclei have no matching cell outline. All are retained. pairs contains 1,540 suggested matches, calculated only when a nonempty nucleus mask lies fully inside exactly one overlapping cell mask. A cell may appear in several matches. These links use full individual masks and are not expert-checked pair labels.

source_mask is the author's combined semantic mask, not a tissue-region map. Its absence for 20250328_CCimage49.png does not remove that image's nucleus, cell or parasite objects: those come from the COCO outlines.

Download and read a sample

from io import BytesIO

import numpy as np
import polars as pl
from huggingface_hub import hf_hub_download
from PIL import Image
from pycocotools import mask as coco_mask

path = hf_hub_download(
    repo_id="csmp-hub/air-leish", filename="train.parquet", repo_type="dataset"
)
row = pl.scan_parquet(path).head(1).collect().to_dicts()[0]


def decode_png(data):
    with Image.open(BytesIO(data)) as image:
        return np.array(image)


image = decode_png(row["image"])  # HWC RGB uint8
nuclei = decode_png(row["inst"])  # HW uint16 instance numbers
cells = decode_png(row["inst_cell"])
parasites = decode_png(row["inst_parasite"])
nucleus_types = decode_png(row["type"])

# Read one complete object mask, including any pixels hidden in the numbered mask.
obj = row["objects"][0]
full_object_mask = coco_mask.decode(
    {"size": list(nuclei.shape), "counts": obj["rle_counts"]}
)

Plot examples

Continue from the loading example to show two Set1 images and one Set2 image.

import matplotlib.pyplot as plt
from skimage.segmentation import find_boundaries

columns = ["id", "source_set", "image", "inst", "inst_cell", "inst_parasite"]
dataset = pl.scan_parquet(path).select(columns)
samples = dataset.filter(pl.col("source_set") == "Set1").head(2).collect().to_dicts()
samples += dataset.filter(pl.col("source_set") == "Set2").head(1).collect().to_dicts()

fig, axes = plt.subplots(3, 4, figsize=(12, 12))
for row_axes, sample in zip(axes, samples):
    rgb = decode_png(sample["image"])
    row_axes[0].imshow(rgb)
    for ax, key, color in zip(
        row_axes[1:],
        ["inst", "inst_cell", "inst_parasite"],
        [(40, 180, 245), (245, 150, 40), (220, 55, 180)],
    ):
        mask = decode_png(sample[key])
        overlay = rgb.astype(np.float32).copy()
        overlay[mask > 0] = 0.55 * overlay[mask > 0] + 0.45 * np.array(color)
        overlay[find_boundaries(mask, mode="inner")] = color
        ax.imshow(overlay.astype(np.uint8))
    row_axes[0].text(0, 1.02, sample["id"], transform=row_axes[0].transAxes, fontsize=9)
    for ax in row_axes:
        ax.axis("off")
for ax, title in zip(axes[0], ["Giemsa image", "Nuclei", "Whole cells", "Parasites"]):
    ax.set_title(title, pad=22)
fig.subplots_adjust(
    left=0.01, right=0.99, bottom=0.01, top=0.94, hspace=0.12, wspace=0.03
)
plt.show()

Two examples without the image underneath. Black marks background; bright instance colours distinguish objects, not cell types. Colours do not indicate nucleus-to-cell matches. The last column enlarges a 256 × 256 region from each parasite mask; no image pixels are shown.

from skimage.color import label2rgb

instance_colors = [
    (1, 0.85, 0.2),
    (0.2, 1, 1),
    (1, 0.3, 0.8),
    (0.5, 1, 0.2),
    (0.65, 0.55, 1),
]
mask_samples = samples
mask_samples = [mask_samples[0], mask_samples[-1]]
mask_layers = [
    ("inst", "Nucleus instances"),
    ("inst_cell", "Whole-cell instances"),
    ("inst_parasite", "Parasite instances"),
    ("inst_parasite", "Parasites: 256 × 256 crop"),
]
mask_fig, mask_axes = plt.subplots(2, 4, figsize=(12.8, 8))
for row_axes, sample in zip(mask_axes, mask_samples):
    for column, (ax, (key, title)) in enumerate(zip(row_axes, mask_layers)):
        labels = decode_png(sample[key])
        if column == 3:
            ys, xs = np.nonzero(labels)
            y, x = int(ys[len(ys) // 2]), int(xs[len(xs) // 2])
            y0 = max(0, min(y - 128, labels.shape[0] - 256))
            x0 = max(0, min(x - 128, labels.shape[1] - 256))
            labels = labels[y0 : y0 + 256, x0 : x0 + 256]
        pixels = label2rgb(
            labels, colors=instance_colors, bg_label=0, bg_color=(0, 0, 0)
        )
        ax.imshow(pixels, interpolation="nearest")
        ax.set_axis_off()
    label = sample["id"]
    row_axes[0].text(0, 1.02, label, transform=row_axes[0].transAxes, fontsize=8)
for ax, (_, title) in zip(mask_axes[0], mask_layers):
    ax.set_title(title, fontsize=10, pad=18)
mask_fig.subplots_adjust(
    left=0.01, right=0.99, bottom=0.02, top=0.94, hspace=0.16, wspace=0.03
)
plt.show()

Preparation and validation

Prepared from AIR_LEISH_dataset_v1.zip in Zenodo record 17384855. The other ZIP in that record is a related earlier collection and is not added as extra images.

Original RGB pixels and polygon coordinates are preserved. Opaque alpha channels were removed. The polygons were filled using the COCO mask rules; no polygon repairs, resizing or new patches were applied. Full object masks are kept separately from the numbered masks. Some Set2 COCO filenames end in .png.png; file_name points to the actual .png file and source_file_name keeps the original text.

Every saved image and every individual object mask was checked against the source. The original combined masks are kept byte-for-byte where supplied. Counts in this archive differ from those in the paper; the table above reports the files used here. manifest.json records file checks and mask details.

Limitations

• These are two laboratory infection models, not a broad tissue or cell-type collection. There are no tissue-region masks or verified physical pixel sizes.
• Some outlines overlap, have irregular geometry or combine nearby nuclei. Some nuclei have no matching cell outline; unmarked areas are not proof that no object is present. Numbered masks cannot show overlapping objects at the same pixel; use the full masks in objects when needed.
• The image audit found 0 exact-duplicate groups, 0 very-similar-image candidates and 236 possible shared-view pairs. Candidate pairs and screening methods are listed in split-audit.json. Two shared-view pairs were confirmed visually and recorded in shared-view-review.json; weaker candidates may reflect fixed optical or background features. Automated screening cannot establish slide or donor independence.
• No official split lists or slide/donor identifiers were found in the release. Keep related views together if creating your own splits, and do not describe an image-level random holdout as a patient-independent test set.

License and citation

The source data are released under CC BY 4.0, as stated in the Zenodo metadata and both COCO annotation files. When sharing this version, credit the creators, link to the source and license, keep the supplied notices and explain your changes. LICENSE contains the full terms and NOTICE.md describes this preparation.

Please cite the data:

Harigua, E.; Oualha, R.; Fekih-Romdhane, N.; Driss, D.; Abdelkrim, Y. Z.; Guizani, I. (2025). AIR-LEISH: An AI-Ready dataset of microscopy images of Leishmania-infected macrophages. Zenodo. doi:10.5281/zenodo.17384855. Institut Pasteur de Tunis.

And the accompanying paper:

Oualha, R.; Fekih-Romdhane, N.; Driss, D.; Abdelkrim, Y. Z.; Guizani, I.; Harigua-Souiai, E. (2026). AIR-LEISH: A Dataset of Giemsa-Stained Microscopy Images for AI-based Leishmania amastigotes Detection. Scientific Data, 13, 328. doi:10.1038/s41597-026-06676-8.