- Rating
- 2
- Confidence
- 4
- Informativeness
- 1
- Presentation
- 2
- Soundness
- 1
Read review
Paper summary
This paper proposes a structured semantic consistency framework for unsupervised domain adaptation (UDA) of nuclei instance segmentation and classification, transferring a multi-head network (NuLite family, FastViT encoder) from a labeled source domain to an unlabeled target domain. The framework combines five components: source foreground feature preservation under appearance changes (SFA); masked target-view consistency, where predictions of a masked view follow confident pseudo-labels from the full view (MIC); alignment of confident target features to source class prototypes (PSA); a loss that matches class-relation matrices between source class centers and fixed CLIP text embeddings (CoPT); and a target class-support constraint that suppresses classes absent from the target label space, used only when label spaces differ (TCC). Training has two stages: joint source-target pre-adaptation, then source-only refinement. The method is evaluated on PanNuke internal transfers (multi-source/multi-target and Breast-centered, with cross-organ comparisons on Breast to Testis, Thyroid, and Bile-duct against UDA baselines) and on adaptation from PanNuke to CoNSeP and Lizard through a unified five-class label space. Class-aware PQ rises from 0.1059 to 0.2335 on CoNSeP and from 0.0733 to 0.1127 on Lizard relative to Source-only, and from 0.4502 to 0.4734 and from 0.4389 to 0.4612 in the two internal settings. Ablations (cumulative addition, component removal, independent addition) are reported, with MIC contributing most. The authors state that target labels are used only for final evaluation and that model selection relies on source validation.
Critical strengths
-
The method does not use target labels in training, and model selection is done with source-domain validation. This keeps the evaluation consistent with the unsupervised setting the paper targets.
-
The motivation of the method is clear, and the components are analyzed in several ways (cumulative addition, removal, and independent addition). The paper also reports results that are unfavorable to the proposed method, such as PSA lowering mPQ in the cumulative sequence. This makes the evidence easier to interpret honestly.
Critical weaknesses
-
The central claim is the combination of existing components, but the evidence that the combination works well is weak. Most components come from prior work: masked-view consistency (MIC, Hoyer et al., 2023), prototype-based alignment (Zhang et al., 2021), and a text-derived class-relation loss described as CoPT-inspired (Mata et al., 2024). Since the integration is presented as the contribution, the paper needs to show that it works well and that the gains are large. It does not. On the external datasets (CoNSeP and Lizard), which carry the headline results, there are no comparison methods, only Source-only. The component analysis is incomplete: the cumulative sequence is run only on Lizard, the independent additions of PSA, CoPT, and TCC give small gains, and MIC alone is never reported, even though removing it drops Lizard mPQ below Source-only. This matters because without this evidence the paper's main contribution cannot be assessed.
-
The comparison with existing methods is hard to trust. In Tables 1-3, the numbers for Source-only, DA-RCNN, MGA, HT, BAFA, and DA-CC appear identical to those reported in Fan et al. (2024), but the paper gives no source and does not say whether the settings (data, backbone, classes) are the same. Fan et al. use Mask R-CNN, whereas this paper uses NuLite, so the gain cannot be attributed to the method rather than to a stronger backbone. The seven methods in Table 1 are listed by name only, without description or citation, and the correspondence with methods in the related work is not stated. This matters because the within-PanNuke comparison is the only place where the method is compared with other adaptation methods.
Minor: Table 9 does not say how the models were trained or evaluated, shows 11 of the 19 PanNuke tissues, and uses inconsistent bold formatting in the Average row. The Dead class is missing from Tables 1-3.
Questions for authors
-
Are the numbers in Tables 1-3 for Source-only, DA-RCNN, MGA, HT, BAFA, and DA-CC taken from Fan et al. (2024) or re-run? If taken from it, are the data, backbone, and classes the same as in your setting? What are DA-CC, DDSPSeg, and FAUDA, and where do the DDSPSeg and FAUDA numbers come from? (Clarification)
-
How were the models in Table 9 trained and evaluated, and why are only 11 of the 19 PanNuke tissues shown? Why is the bPQ gain in Table 9 (about 0.051) more than twice the overall gain in Table 8 (about 0.021)? (Clarification)
-
Could you report Source-only + MIC, and the removal of SFA and of TCC? Without them it is hard to tell how much each component contributes. (Suggestion)
-
The reported Source-only on Lizard (bPQ 0.2182, mPQ 0.0733) looks low. In my own experiments with existing state-of-the-art fully-supervised nuclei segmentation and classification models (CellViT, PromptNucSeg, ...) trained on PanNuke and applied directly to Lizard, the PQ was considerably higher. Did you account for the magnification difference between PanNuke and Lizard? Could you report how such fully-supervised models perform on Lizard as a reference? (Clarification and suggestion)