SegPC-2021 — Segmentation of Multiple Myeloma Plasma Cells

Public mirror of the ISBI-2021 SegPC challenge dataset (instance segmentation of Multiple Myeloma plasma cells in microscopic images), created by SBILab, IIIT-Delhi & AIIMS New Delhi (Anubha Gupta, Ritu Gupta et al.). Mirrored for the EasyMedSeg / MedSphere segmentation benchmark.

What this is

Brightfield microscopy of bone-marrow aspirate slides from Multiple Myeloma patients, Jenner–Giemsa stained and stain-color-normalized (GCTI-SN). The task is to segment each cell of interest (cytoplasm + nucleus).

Splits (only ground-truth-bearing data is mirrored)

split	images	notes
train	298	full train set, all with GT
validation	199	source has 200 images; 610.bmp ships no mask and is dropped

The challenge test split (277 images) is excluded — its ground truth is withheld by the organizers (test was leaderboard-only).

Acquisition / image sizes

Two cameras were used: Olympus (2040×1536) and Nikon (2560×1920). The Nikon images are flagged is_mimm_subset = true — see the overlap warning below.

Columns

• image — RGB microscopy image (PNG, pixel-lossless re-encode of the source BMP).
• mask — semantic ground truth, {0 = background, 1 = cytoplasm, 2 = nucleus}.
  • whole-cell GT = mask > 0; nucleus GT = mask == 2; cytoplasm-only = mask == 1.
• instance_mask — {0 = background, 1..N}, a distinct id per cell of interest.
• overlay — half-resolution viewer aid (nucleus=red, cytoplasm=green); not ground truth.
• image_id, split, num_instances, height, width, is_mimm_subset, acquisition.

Mask encoding note (reformatting)

The original release ships one grayscale BMP per cell instance (y/<id>_<k>.bmp) with raw values {0, 20, 40} (background / cytoplasm / nucleus). This mirror reformats those into the per-image mask (semantic 0/1/2) and instance_mask (instance ids) columns above. No pixels are otherwise altered.

⚠️ "Cells of interest" caveat

Ground truth is provided only for designated cells of interest — a microscopic field may contain additional plasma cells that are intentionally left unlabeled (treated as background). The official challenge metric evaluates cells of interest only.

⚠️ Cross-dataset overlap (evaluation leakage)

The Nikon 2560×1920 images (is_mimm_subset = true) correspond to the MiMM_SBILab dataset, and SegPC shares the same SBILab/AIIMS Jenner-Giemsa MM pipeline as the SN-AM dataset. No per-image cross-reference table is published; dedup by image content before co-benchmarking SegPC-2021 against MiMM_SBILab or SN-AM(MM).

License

CC BY-NC-SA 4.0 — non-commercial, share-alike, attribution required. By using this data you agree to cite the publications below.

Citation

1. A. Gupta, R. Duggal, S. Gehlot, R. Gupta, A. Mangal, L. Kumar, N. Thakkar, D. Satpathy, "GCTI-SN: Geometry-Inspired Chemical and Tissue Invariant Stain Normalization of Microscopic Medical Images," Medical Image Analysis 65, 2020. doi:10.1016/j.media.2020.101788
2. S. Gehlot, A. Gupta, R. Gupta, "EDNFC-Net: Convolutional Neural Network with Nested Feature Concatenation for Nuclei-Instance Segmentation," ICASSP 2020, pp. 1389–1393.
3. A. Gupta, P. Mallick, O. Sharma, R. Gupta, R. Duggal, "PCSeg: Color model driven probabilistic multiphase level set based tool for plasma cell segmentation in multiple myeloma," PLoS ONE 13(12): e0207908, 2018. doi:10.1371/journal.pone.0207908
4. A. Gupta, S. Gehlot, S. Goswami, et al., "SegPC-2021: A challenge & dataset on segmentation of Multiple Myeloma plasma cells from microscopic images," Medical Image Analysis 83, 2023. doi:10.1016/j.media.2022.102677

Dataset DOI: 10.21227/7np1-2q42 (IEEE DataPort). Source: Kaggle sbilab/segpc2021dataset.