LyNSeC - Lymphoma Nuclear Segmentation and Classification

Histopathology dataset of diffuse large B-cell lymphoma (DLBCL) tissue with per-nucleus instance segmentation masks across both H&E and IHC stains. Released alongside the HoLy-Net paper (Naji et al., Comput Biol Med 2024).

Source: https://zenodo.org/records/8065174 Paper: https://www.sciencedirect.com/science/article/pii/S0010482524000623

Overview

• Modality: Histopathology (H&E + IHC brightfield microscopy)
• Tissue: Lymph-node tissue, DLBCL
• Stains: H&E, plus IHC markers CD3 / Ki67 / ERG
• Tile spec: 512x512 RGB, 40x magnification
• Total samples: 699 tiles, 161,247 annotated nuclei

Splits

Split	Stain	Tiles	Nuclei	Notes
lynsec1_ihc	IHC (CD3 / Ki67 / ERG)	379	87,316	Per-nucleus marker-positive vs marker-negative class label
lynsec2_he	H&E	280	65,479	Instance-only (no per-cell class label)
lynsec3_he_expert	H&E (expert-refined subset of LyNSeC 2)	40	8,452	Per-nucleus tumor vs non-tumor class label

The Zenodo release does NOT ship predefined train/val/test split files; splits in the original paper are defined in code, not as released folders.

Columns

Column	Type	Notes
id	string	Tile identifier from the Zenodo release (e.g. 0_l1)
image	Image (RGB)	512x512 RGB tile
instance_mask	Image (16-bit grayscale PNG, mode I;16)	Per-nucleus instance ID; 0 = background, 1..N = instance IDs (N up to 405 in lynsec2_he)
class_mask	Image (mode L, 8-bit grayscale) - only on lynsec1_ihc and lynsec3_he_expert	Per-pixel class label: 0 = background. Non-zero class semantics are inherited from the source paper; see notes below
num_nuclei	int32	Number of instances in the tile (= instance_mask.max())

Class label semantics

• lynsec1_ihc: classes {0, 1, 2}. Per the source paper, class indicates marker positivity (one of marker-negative / marker-positive). The Zenodo release does not document the specific mapping of 1 and 2 to those two states - downstream users are expected to verify against the source code.
• lynsec3_he_expert: classes {0, 1, 2}. Per the source paper, class indicates tumor vs non-tumor cells. Same caveat about which integer maps to which class.

For binary semantic segmentation (nucleus vs background) treat instance_mask > 0 as foreground.

Derivation

Each Zenodo .npy ships as (512, 512, C) int32 with C in {4, 5}:

• channels 0..2 = RGB image (values 0..255 stored as int32),
• channel 3 = instance map,
• channel 4 = class map (only when C == 5).

This release re-encodes each tile as:

• image = Image.fromarray(arr[:,:,:3].astype(np.uint8), 'RGB') -> PNG
• instance_mask = Image.fromarray(arr[:,:,3].astype(np.uint16), 'I;16') -> 16-bit PNG
• class_mask = Image.fromarray(arr[:,:,4].astype(np.uint8), 'L') -> 8-bit PNG (lynsec1/3 only)

No other preprocessing.

License

CC BY 4.0 (per Zenodo record 8065174).

Citation

@article{naji2024holynet,
  title   = {HoLy-Net: Segmentation of histological images of diffuse large
             B-cell lymphoma},
  author  = {Naji, Hussein and Sancere, Lucas and Simon, Adrian and Buttner,
             Reinhard and Eich, Marie-Lisa and Lohneis, Philipp and Bozek,
             Katarzyna},
  journal = {Computers in Biology and Medicine},
  volume  = {170},
  pages   = {107978},
  year    = {2024},
  doi     = {10.1016/j.compbiomed.2024.107978}
}