Larrivhan/KidNet-dataset

KidNet Renal Injury Pathology Dataset

KidNet is a curated hematoxylin and eosin (H&E) kidney pathology image dataset for renal injury recognition. Each sample contains one microscopy image and one LabelMe annotation file with polygon-level labels for renal tubules and related pathological structures.

The dataset was built for small-sample renal injury modeling, especially:

• pixel-level segmentation of injury_tubules
• tile-level injury classification
• weakly supervised heatmap localization
• family-level generalization analysis across whole-slide-image (WSI) groups

Dataset Summary

Item	Count
Images	411
Annotation files	411
WSI families	11
Injury-positive images	319
Injury-negative images	92
Total annotated shapes	16,932
injury_tubules annotations	8,996

Directory Structure

Each sample is stored in its own folder:

KidNet/
  WSI1_1/
    WSI1_1.jpg
    WSI1_1.json
  WSI1_2/
    WSI1_2.jpg
    WSI1_2.json
  ...

Each .json file follows the LabelMe format and contains:

• imagePath: image filename
• imageHeight, imageWidth: original image size
• shapes: polygon or circle annotations
• label: annotation class name
• points: vertex coordinates in image pixel space
• shape_type: usually polygon, with occasional circle

Some LabelMe files may contain an imageData field. The paired .jpg file is the authoritative image file; imageData can be removed before upload if a smaller repository size is required.

Label Schema

Label	Count	Description
injury_tubules	8,996	Tubules annotated as injured; main binary segmentation target
healthy_tubules	3,756	Tubules annotated as morphologically healthy
necrotic_tubules	2,369	Necrotic tubule regions
cast	688	Tubular cast regions
glomerulus	665	Glomerular structures
unknown	458	Ambiguous or uncertain regions

For binary renal injury segmentation, use injury_tubules as the positive class and all other pixels as background. For broader pathology modeling, the remaining labels can be used as auxiliary or multilabel targets.

Family-Level Distribution

The recommended split unit is the WSI family, which is the prefix of each sample ID before the final numeric index. Random image-level splitting is not recommended because images from the same family may share staining, acquisition, tissue-source, and morphology patterns.

Family	Images	Injury-positive	Injury-negative	Injury annotations	Total shapes
WSI1	20	2	18	5	669
WSI14	10	0	10	0	380
WSI15	10	0	10	0	383
WSI19	103	103	0	2,046	2,766
WSI2	20	5	15	18	720
WSI20	36	36	0	1,242	1,989
WSI3	34	20	14	770	1,658
WSI4	20	19	1	168	946
WSI5	118	115	3	4,587	5,780
WSI6	20	8	12	61	907
WSI7	20	11	9	99	734

Recommended Evaluation Protocol

Use family-level held-out evaluation:

1. Select one WSI family as the test family.
2. Select another WSI family as validation.
3. Train on the remaining families.
4. Repeat across all 11 held-out families.

This protocol is stricter than random image splitting and better measures generalization to unseen WSI families.

Recommended metrics:

• Segmentation: Dice, IoU, precision, recall, specificity
• Tile classification: recall, precision, F1, balanced accuracy, AUROC when applicable
• Heatmap localization: image-level recall, false-positive area, thresholded heatmap quality

Loading Example

from pathlib import Path
import json
from PIL import Image

root = Path("KidNet")
samples = []

for sample_dir in sorted(p for p in root.iterdir() if p.is_dir()):
    image_path = next(sample_dir.glob("*.jpg"))
    json_path = next(sample_dir.glob("*.json"))
    with json_path.open("r", encoding="utf-8") as f:
        ann = json.load(f)

    labels = [shape["label"] for shape in ann.get("shapes", [])]
    samples.append(
        {
            "sample_id": image_path.stem,
            "family": image_path.stem.split("_")[0],
            "image": Image.open(image_path).convert("RGB"),
            "annotation": ann,
            "has_injury": "injury_tubules" in labels,
        }
    )

print(len(samples))

Converting injury_tubules to a Binary Mask

from PIL import Image, ImageDraw

def injury_mask(annotation):
    width = int(annotation["imageWidth"])
    height = int(annotation["imageHeight"])
    mask = Image.new("L", (width, height), 0)
    draw = ImageDraw.Draw(mask)

    for shape in annotation.get("shapes", []):
        if shape.get("label") != "injury_tubules":
            continue
        points = [tuple(p) for p in shape.get("points", [])]
        if shape.get("shape_type") == "polygon" and len(points) >= 3:
            draw.polygon(points, fill=1)
        elif shape.get("shape_type") == "circle" and len(points) >= 2:
            (cx, cy), (px, py) = points[:2]
            r = ((px - cx) ** 2 + (py - cy) ** 2) ** 0.5
            draw.ellipse((cx - r, cy - r, cx + r, cy + r), fill=1)

    return mask

Intended Use

This dataset is intended for academic research on renal injury recognition from pathology images. Suitable use cases include segmentation baselines, weakly supervised classification, heatmap localization, and small-sample generalization studies.

The dataset is not intended for clinical diagnosis, treatment decisions, or deployment as a medical device.

Limitations

• The dataset is small and strongly imbalanced across WSI families.
• Some families are injury-rich, while others are sparse-positive or fully negative.
• Labels are research annotations and should not be treated as exhaustive clinical ground truth.
• Pixel-level boundaries can be uncertain for subtle tubular injury patterns.
• Models evaluated with random image-level splits may report overly optimistic performance.

Ethics And Privacy

The current release contains experimental kidney histopathology images and does not include human-identifiable personal information. Users should still follow institutional, animal research, and data-use requirements applicable to their own setting.

License

License information should be confirmed by the dataset owner before public redistribution. The current metadata uses other as a placeholder. If the dataset is released publicly, replace it with the final approved license.

Citation

If you use this dataset, please cite the project or competition report associated with KidNet. A formal citation can be added here after release.

@dataset{kidnet_renal_injury_pathology,
  title = {KidNet Renal Injury Pathology Dataset},
  year = {2026},
  note = {H&E kidney pathology images with LabelMe annotations for renal injury recognition}
}