ImPartial: Multi-channel Whole-Cell Segmentation using Partial Annotations

Gunjan Shrivastava; Saad Nadeem

arxiv: 2605.24128 · v1 · pith:UPC53SFKnew · submitted 2026-05-22 · 💻 cs.CV

ImPartial: Multi-channel Whole-Cell Segmentation using Partial Annotations

Gunjan Shrivastava , Saad Nadeem This is my paper

Pith reviewed 2026-06-30 15:28 UTC · model grok-4.3

classification 💻 cs.CV

keywords cell segmentationpartial annotationsself-supervised learningmultiplexed imagingimmunohistochemistrywhole-cell segmentationdeep learningscribble supervision

0 comments

The pith

ImPartial reaches fully supervised cell segmentation accuracy using only sparse scribble annotations.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper presents ImPartial as a framework that augments a segmentation network with self-supervised multi-channel quantized imputation. This addition lets the model learn from partial scribble labels on multi-channel pathology images without needing dense pixel annotations. The approach rests on the idea that exact image reconstruction is unnecessary and that a classification-style self-supervised task better supports the segmentation objective. Experiments on multiplexed cellular datasets and brightfield immunohistochemistry images show performance matching fully supervised baselines while using far fewer labels.

Core claim

ImPartial combines partial scribble supervision with self-supervised multi-channel quantized imputation to produce whole-cell segmentations whose accuracy matches that of models trained on complete pixel-wise labels. The imputation step converts reconstruction into a quantized classification task across channels, which the authors show aligns more directly with segmentation than pixel-level denoising would.

What carries the argument

Self-supervised multi-channel quantized imputation, which augments the segmentation loss by turning partial image reconstruction into a per-channel classification objective.

If this is right

Multiplexed imaging datasets with arbitrary channel counts become trainable with only sparse scribbles instead of exhaustive masks.
Clinical brightfield immunohistochemistry slides can be segmented at scale without requiring dense expert labeling.
Annotation effort for emerging biological modalities drops while maintaining parity with full-supervision baselines.
The same partial-label plus imputation recipe applies across both multiplexed fluorescence and single-plex brightfield domains.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

The method suggests that many dense-labeling tasks in imaging could be reframed around classification-style self-supervision rather than reconstruction.
If the quantized imputation generalizes, it may reduce the need for modality-specific full-supervision pipelines in pathology.
One could test whether the same partial-annotation regime works for 3D volumetric cell segmentation without additional architectural changes.

Load-bearing premise

The premise that perfect pixel-wise image reconstruction is unnecessary and that a classification-based self-supervised task aligns better with segmentation than standard reconstruction losses.

What would settle it

An ablation on the same partial-annotation training sets in which the quantized imputation head is removed or replaced by a standard pixel-reconstruction loss and the resulting segmentation accuracy falls below that of a fully supervised model trained on full labels.

Figures

Figures reproduced from arXiv: 2605.24128 by Gunjan Shrivastava, Saad Nadeem.

**Figure 2.** Figure 2: Visual comparison on TissueNet dataset results with ground truth, under [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗

**Figure 3.** Figure 3: (Top) Quantitative results on 6-channel CPDMFCI dataset with ground [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗

**Figure 4.** Figure 4: Qualitative results (top) on IHC dataset and metrics performance com [PITH_FULL_IMAGE:figures/full_fig_p008_4.png] view at source ↗

read the original abstract

Accurate cell segmentation in pathology images typically requires dense pixel-wise annotations, which are costly and time-consuming to obtain. This challenge is especially important for emerging biological imaging modalities and multiplexed datasets with variable channel configurations, where expert-labeled data are scarce. In this work, we introduce ImPartial, a deep learning framework designed to achieve state-of-the-art segmentation performance in low-annotation regimes using sparse scribbles and limited supervision. ImPartial augments the segmentation objective via self-supervised multi-channel quantized imputation. This approach leverages the observation that perfect pixel-wise reconstruction or denoising of the image is not needed for accurate segmentation, and thus, introduces a self-supervised classification objective that better aligns with the overall segmentation goal. We demonstrate that ImPartial achieves performance at par with fully supervised models while requiring substantially fewer annotations. Extensive experiments on benchmark multiplexed cellular imaging and single-plex clinical brightfield immunohistochemistry datasets show consistent improvements over strong baselines with only partial annotations. All benchmark datasets and code are available via our Github: https://github.com/nadeemlab/ImPartial.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

ImPartial pairs scribble supervision with a quantized multi-channel imputation auxiliary task and claims near-parity with full supervision, but the ablation isolating the classification loss is absent.

read the letter

The main point is that this work takes partial scribble labels for whole-cell segmentation in multiplexed pathology images and adds a self-supervised task that quantizes the channels and trains an imputation classifier instead of a reconstruction head. They argue this aligns better with the segmentation objective because perfect pixel recovery is unnecessary. Experiments on multiplexed cellular and brightfield IHC benchmarks show gains over baselines with far fewer annotations, and the code plus data are released.

The approach is practical for modalities where dense labels are expensive. The release of everything makes it straightforward to check.

The soft spot is exactly the one flagged in the stress test: the motivation hinges on the classification formulation being responsible for the gains, yet there is no ablation that runs the identical architecture and partial-label pipeline once with the quantized loss and once with MSE reconstruction. Without that comparison it is possible the benefit comes from the multi-channel structure or the scribble loss alone. The abstract also gives no numbers, so the full paper must supply the metrics and controls to back the parity claim.

This is for labs doing segmentation in emerging multiplexed or low-annotation settings. A reader already working on semi-supervised medical imaging would find the auxiliary-task design worth looking at.

It deserves peer review because the problem is concrete and the method is reproducible even if the central design choice needs tighter evidence.

Referee Report

2 major / 1 minor

Summary. The paper introduces ImPartial, a framework for multi-channel whole-cell segmentation in pathology images that combines sparse scribble annotations with a self-supervised multi-channel quantized imputation task. It claims this yields segmentation performance at parity with fully supervised models while requiring substantially fewer annotations, with consistent improvements over baselines on multiplexed cellular imaging and brightfield IHC datasets.

Significance. If the central claim is supported by the experiments, the work would be significant for reducing annotation costs in low-supervision regimes for multiplexed and emerging imaging modalities. The release of code and benchmarks strengthens reproducibility.

major comments (2)

[Methods / Experiments] The abstract and introduction motivate the quantized classification objective by claiming that 'perfect pixel-wise reconstruction or denoising of the image is not needed' and that a classification objective 'better aligns' with segmentation. However, no ablation is described that trains the identical architecture and partial-annotation pipeline with the proposed classification loss versus a standard regression (MSE) reconstruction loss. This comparison is required to establish that the alignment argument, rather than the multi-channel structure or scribble supervision alone, drives the reported gains.
[Abstract / §4 Experiments] The central claim of performance parity with fully supervised models is stated without quantitative support in the abstract (no Dice/IoU values, no baseline tables, no annotation-fraction curves). The full experimental section must provide these metrics with error bars and statistical tests to substantiate parity under substantially reduced annotation budgets.

minor comments (1)

[Abstract] The GitHub link is provided but the manuscript should specify the exact commit or release tag used for the reported results to ensure reproducibility.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive feedback. We address the two major comments point-by-point below and will revise the manuscript to incorporate the suggested improvements.

read point-by-point responses

Referee: [Methods / Experiments] The abstract and introduction motivate the quantized classification objective by claiming that 'perfect pixel-wise reconstruction or denoising of the image is not needed' and that a classification objective 'better aligns' with segmentation. However, no ablation is described that trains the identical architecture and partial-annotation pipeline with the proposed classification loss versus a standard regression (MSE) reconstruction loss. This comparison is required to establish that the alignment argument, rather than the multi-channel structure or scribble supervision alone, drives the reported gains.

Authors: We agree that a direct ablation comparing the quantized classification loss to an MSE reconstruction loss on the same architecture and partial-annotation pipeline would strengthen the justification for our design choice. The current experiments demonstrate the overall framework but do not isolate this factor. We will add this ablation study to the revised manuscript. revision: yes
Referee: [Abstract / §4 Experiments] The central claim of performance parity with fully supervised models is stated without quantitative support in the abstract (no Dice/IoU values, no baseline tables, no annotation-fraction curves). The full experimental section must provide these metrics with error bars and statistical tests to substantiate parity under substantially reduced annotation budgets.

Authors: The experimental section already reports Dice/IoU metrics, baseline comparisons, and annotation-fraction results across the multiplexed and IHC datasets. However, the abstract contains only qualitative claims. We will revise the abstract to include key quantitative results and will ensure error bars plus statistical tests are explicitly presented or added in §4 of the revised version. revision: partial

Circularity Check

0 steps flagged

No circularity; self-supervised objective is independent design choice with no derivation or self-citation reduction

full rationale

The paper presents ImPartial as a framework that augments segmentation with a self-supervised multi-channel quantized imputation task. This is motivated by an explicit observation ('perfect pixel-wise reconstruction or denoising of the image is not needed for accurate segmentation, and thus, introduces a self-supervised classification objective that better aligns with the overall segmentation goal') but is not derived from or fitted to the target labels. No equations, uniqueness theorems, or self-citations are invoked to force the choice; the component is described as an independent augmentation. Central claims rest on empirical benchmarks rather than any closed derivation chain that reduces to inputs by construction. This matches the default case of a self-contained empirical method.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The central claim rests on the domain assumption that a quantized imputation classification objective aligns with segmentation without needing full reconstruction; no free parameters or invented entities are mentioned in the abstract.

axioms (1)

domain assumption Perfect pixel-wise reconstruction or denoising of the image is not needed for accurate segmentation
Explicitly stated as the observation leveraged by the method.

pith-pipeline@v0.9.1-grok · 5707 in / 1032 out tokens · 33331 ms · 2026-06-30T15:28:49.736810+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

12 extracted references · 1 canonical work pages

[1]

Scientific Data10(1), 193 (2023)

Aleynick, N., Li, Y., Xie, Y., Zhang, M., Posner, A., Roshal, L., Pe’er, D., Van- guri, R.S., Hollmann, T.J.: Cross-platform dataset of multiplex fluorescent cellular object image annotations. Scientific Data10(1), 193 (2023)

2023
[2]

European Conference on Computer Vision pp

Buchholz, T.O., Prakash, M., Krull, A., Jug, F.: Denoiseg: Joint denoising and segmentation. European Conference on Computer Vision pp. 324–337 (2020)

2020
[3]

Nature Machine Intelligence4(4), 401–412 (2022)

Ghahremani, P., Li, Y., Kaufman, A., Vanguri, R., Greenwald, N., Angelo, M., Hollmann, T.J., Nadeem, S.: Deep learning-inferred multiplex immunofluorescence for immunohistochemical image quantification. Nature Machine Intelligence4(4), 401–412 (2022)

2022
[4]

Nature biotechnology40(4), 555–565 (2022)

Greenwald, N.F., Miller, G., Moen, E., Kong, A., Kagel, A., Dougherty, T., Full- away, C.C., McIntosh, B.J., Leow, K.X., Schwartz, M.S., et al.: Whole-cell seg- mentation of tissue images with human-level performance using large-scale data annotation and deep learning. Nature biotechnology40(4), 555–565 (2022)

2022
[5]

In: Proceedings of the IEEE/CVF International Conference on Computer Vision

Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., et al.: Segment anything. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 4015–4026 (2023)

2023
[6]

In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition

Krull, A., Buchholz, T.O., Jug, F.: Noise2void-learning denoising from single noisy images. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 2129–2137 (2019)

2019
[7]

Nature Methods22(12), 2585–2593 (Dec 2025)

Marks, M., Israel, U., Dilip, R., Li, Q., Yu, C., Laubscher, E., Iqbal, A., Pradhan, E., Ates, A., Abt, M., Brown, C., Pao, E., Li, S., Pearson-Goulart, A., Perona, P., Gkioxari, G., Barnowski, R., Yue, Y., Van Valen, D.: CellSAM: a foundation model for cell segmentation. Nature Methods22(12), 2585–2593 (Dec 2025)

2025
[8]

Nature methods19(12), 1634–1641 (2022)

Pachitariu, M., Stringer, C.: Cellpose 2.0: how to train your own model. Nature methods19(12), 1634–1641 (2022)

2022
[9]

Nature Methods (2024)

Pang, M., Roy, T.K., Wu, X., Tan, K.: Cellotype: A unified model for segmentation and classification of tissue images. Nature Methods (2024)

2024
[10]

In: The IEEE Winter Conference on Applications of Computer Vision (WACV) (March 2020)

Weigert, M., Schmidt, U., Haase, R., Sugawara, K., Myers, G.: Star-convex poly- hedra for 3d object detection and segmentation in microscopy. In: The IEEE Winter Conference on Applications of Computer Vision (WACV) (March 2020). https://doi.org/10.1109/WACV45572.2020.9093435

work page doi:10.1109/wacv45572.2020.9093435 2020
[11]

In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision

Wen, R., Yuan, H., Ni, D., Xiao, W., Wu, Y.: From denoising training to test-time adaptation: Enhancing domain generalization for medical image segmentation. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 464–474 (2024)

2024
[12]

ACM Transactions on Multimedia Computing, Communications and Applications19(2), 1–23 (2023)

Xu, S., Sun, K., Liu, D., Xiong, Z., Zha, Z.J.: Synergy between semantic segmenta- tion and image denoising via alternate boosting. ACM Transactions on Multimedia Computing, Communications and Applications19(2), 1–23 (2023)

2023

[1] [1]

Scientific Data10(1), 193 (2023)

Aleynick, N., Li, Y., Xie, Y., Zhang, M., Posner, A., Roshal, L., Pe’er, D., Van- guri, R.S., Hollmann, T.J.: Cross-platform dataset of multiplex fluorescent cellular object image annotations. Scientific Data10(1), 193 (2023)

2023

[2] [2]

European Conference on Computer Vision pp

Buchholz, T.O., Prakash, M., Krull, A., Jug, F.: Denoiseg: Joint denoising and segmentation. European Conference on Computer Vision pp. 324–337 (2020)

2020

[3] [3]

Nature Machine Intelligence4(4), 401–412 (2022)

Ghahremani, P., Li, Y., Kaufman, A., Vanguri, R., Greenwald, N., Angelo, M., Hollmann, T.J., Nadeem, S.: Deep learning-inferred multiplex immunofluorescence for immunohistochemical image quantification. Nature Machine Intelligence4(4), 401–412 (2022)

2022

[4] [4]

Nature biotechnology40(4), 555–565 (2022)

Greenwald, N.F., Miller, G., Moen, E., Kong, A., Kagel, A., Dougherty, T., Full- away, C.C., McIntosh, B.J., Leow, K.X., Schwartz, M.S., et al.: Whole-cell seg- mentation of tissue images with human-level performance using large-scale data annotation and deep learning. Nature biotechnology40(4), 555–565 (2022)

2022

[5] [5]

In: Proceedings of the IEEE/CVF International Conference on Computer Vision

Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., et al.: Segment anything. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 4015–4026 (2023)

2023

[6] [6]

In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition

Krull, A., Buchholz, T.O., Jug, F.: Noise2void-learning denoising from single noisy images. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 2129–2137 (2019)

2019

[7] [7]

Nature Methods22(12), 2585–2593 (Dec 2025)

Marks, M., Israel, U., Dilip, R., Li, Q., Yu, C., Laubscher, E., Iqbal, A., Pradhan, E., Ates, A., Abt, M., Brown, C., Pao, E., Li, S., Pearson-Goulart, A., Perona, P., Gkioxari, G., Barnowski, R., Yue, Y., Van Valen, D.: CellSAM: a foundation model for cell segmentation. Nature Methods22(12), 2585–2593 (Dec 2025)

2025

[8] [8]

Nature methods19(12), 1634–1641 (2022)

Pachitariu, M., Stringer, C.: Cellpose 2.0: how to train your own model. Nature methods19(12), 1634–1641 (2022)

2022

[9] [9]

Nature Methods (2024)

Pang, M., Roy, T.K., Wu, X., Tan, K.: Cellotype: A unified model for segmentation and classification of tissue images. Nature Methods (2024)

2024

[10] [10]

In: The IEEE Winter Conference on Applications of Computer Vision (WACV) (March 2020)

Weigert, M., Schmidt, U., Haase, R., Sugawara, K., Myers, G.: Star-convex poly- hedra for 3d object detection and segmentation in microscopy. In: The IEEE Winter Conference on Applications of Computer Vision (WACV) (March 2020). https://doi.org/10.1109/WACV45572.2020.9093435

work page doi:10.1109/wacv45572.2020.9093435 2020

[11] [11]

In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision

Wen, R., Yuan, H., Ni, D., Xiao, W., Wu, Y.: From denoising training to test-time adaptation: Enhancing domain generalization for medical image segmentation. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 464–474 (2024)

2024

[12] [12]

ACM Transactions on Multimedia Computing, Communications and Applications19(2), 1–23 (2023)

Xu, S., Sun, K., Liu, D., Xiong, Z., Zha, Z.J.: Synergy between semantic segmenta- tion and image denoising via alternate boosting. ACM Transactions on Multimedia Computing, Communications and Applications19(2), 1–23 (2023)

2023