DeepLIIF (Deep Learning-Inferred Immunofluorescence)

Co-registered IHC (Ki67-DAB brightfield) and multiplex immunofluorescence (mpIF) patches with cell-level segmentation + classification ground truth for quantification of clinical pathology slides.

Source paper: Ghahremani et al., "Deep learning-inferred multiplex immunofluorescence for immunohistochemical image quantification," Nature Machine Intelligence 4(4):401-412, 2022. DOI: 10.1038/s42256-022-00471-x. Zenodo: https://zenodo.org/records/4751737.

Overview

• Modality: Histopathology - IHC brightfield + co-registered mpIF (DAPI, Lap2, Ki67 marker)
• Patch size: 512x512 each (originals are 3072x512 PNGs concatenating six panels)
• Tissue: Bladder + Lung (main DeepLIIF), Breast carcinoma (BC-DeepLIIF subset)
• Total samples: 1,715
  • train (DeepLIIF): 575
  • validation (DeepLIIF): 91
  • test (DeepLIIF): 598
  • bc_train (BC-DeepLIIF, breast): 385
  • bc_validation (BC-DeepLIIF, breast): 66

Columns

Column	Type	Notes
sample_id	string	Original PNG stem
tissue	ClassLabel(3)	0=BC, 1=Bladder, 2=Lung
subset	string	DeepLIIF (main) or BC-DeepLIIF
ihc	Image (RGB)	Input - 512x512 brightfield Ki67-DAB IHC
hematoxylin	Image (RGB)	Aux target - reconstructed hematoxylin channel
dapi	Image (RGB)	Aux target - mpIF DAPI nuclear stain
lap2	Image (RGB)	Aux target - mpIF Lap2 nuclear-envelope stain
marker	Image (RGB)	Aux target - mpIF Ki67 marker channel
seg_mask	Image (RGB)	Ground truth - red=Ki67+ cell, blue=Ki67- cell, green=boundary, black=background

Ground Truth

The seg_mask column is the canonical GT. It was generated by combining per-modality instance segmentations from mpIF DAPI + Lap2 + Hematoxylin + IHC, with instance boundaries initialized by the ImPartial interactive framework on DAPI. Cells are then classified red/blue based on Ki67 marker positivity.

For binary semantic segmentation (nucleus vs background) treat (red OR blue) pixels as foreground. For positive-vs-negative classification, decode red and blue channels separately.

Derivation

Each source PNG was sliced column-wise into six 512x512 panels:

• Columns [0:512] -> ihc
• Columns [512:1024] -> hematoxylin
• Columns [1024:1536] -> dapi
• Columns [1536:2048] -> lap2
• Columns [2048:2560] -> marker
• Columns [2560:3072] -> seg_mask

No other preprocessing.

License

CC BY 4.0 (dataset, per Zenodo record 4751737). The DeepLIIF code repo is Apache 2.0 with Commons Clause; that license applies only to code/models, not to this imaging data.

Citation

@article{ghahremani2022deep,
  title={Deep learning-inferred multiplex immunofluorescence for
         immunohistochemical image quantification},
  author={Ghahremani, Parmida and Li, Yanyun and Kaufman, Arie and
          Vanguri, Rami and Greenwald, Noah and Angelo, Michael and
          Hollmann, Travis J and Nadeem, Saad},
  journal={Nature Machine Intelligence},
  volume={4}, number={4}, pages={401--412}, year={2022},
  doi={10.1038/s42256-022-00471-x}
}