Pathologists can click the center of a cell nucleus in about a second, but tracing its outline takes far longer. Click-supervised methods therefore learn where nuclei are and leave open where they end. The open question matters most where it is hardest: in frozen or poorly stained tissue the membrane fades into the surrounding stroma, and nuclei press against each other. Current click-trained systems handle this in one of three ways. They draw fixed geometric partitions between clicks, they trust low-level color edges that are absent in these images, or they group pixel features with a clustering step that takes seconds per image.
We propose a different source of boundary supervision that clicks already contain but that no current method uses as a training signal. If two clicks mark two different nuclei, then walking from one click toward the other, ownership of the pixels must hand over from the first nucleus to the second exactly once. Walking outward from any click, membership in that nucleus can only fall, never rise again. These ordering facts fix the shape of each nucleus's mask profile without fixing where the edge falls. A single-stage network, meaning one that predicts every instance mask in one forward pass with no proposal-then-crop step, can then learn the edge location from its own features. A contrastive term rewards placing the crossing where foreground and background features separate. At test time there is no clustering and no post-processing.
Core Claim
Click-only training can supply a real boundary loss to a single-stage instance segmenter, rather than an edge-blind region loss on heuristic pseudo-masks. The supervision is a set of order constraints along segments between neighboring clicks and along rays from each click. These are paired with a duality term that ties a boundary head to the spatial gradient of the predicted masks. We claim this beats fixed Voronoi partitioning and color-affinity boundary cues specifically on touching, low-contrast nuclei, and runs at single-stage speed.
The novelty is narrow and testable. The clicks never generate a mask. They generate inequalities the mask must satisfy, and the network chooses the edge.
Research Program
The network follows the keypoint-plus-dynamic-filter design of PointNu-Net. That model is fully supervised: it predicts nucleus-center heatmaps and, for each detected center, generates a small convolution filter that is applied to high-resolution features to produce that nucleus's mask. Clicks supervise its heatmap head directly at no extra cost. The open question is how to supervise the mask filters without masks.
Let $m_i(x) \in [0,1]$ be the predicted soft mask of the instance at click $p_i$. Let $N(i)$ be the click's neighbors in the Delaunay triangulation, meaning the pairs of clicks with no third click between them. Sample points $x_t = p_i + t(p_j - p_i)$ for $t \in [0,1]$. The handover loss is
$$ \mathcal{L}_{\text{order}} = \sum_{i,\, j \in N(i)} \sum_{t} \Big[ \max\big(0,\, m_i(x_{t+\delta}) - m_i(x_t)\big) + \max\big(0,\, m_i(x_t) + m_j(x_t) - 1\big) \Big], $$
with anchors $m_i(p_i)=1$ and $m_i(p_j)=0$, where $\delta$ is the sampling step. The first term forbids $m_i$ from rising as one walks toward the neighbor, and the second forbids two nuclei from claiming the same pixel. The same monotonicity term is applied along $K$ rays from each click. Along those rays the background anchor is a radius $R$ set from local click density, beyond which pixels are treated as background unless claimed. This encodes the roughly star-shaped geometry of nuclei.
These constraints leave the crossing point free. Two terms place it. The first is a contrastive loss in the style of BASS, which samples pixels just inside and just outside a boundary and pushes their features apart. BASS defines those bands with a fixed 4-pixel offset from dense or semi-supervised masks. We instead sample at offsets proportional to the distance from the click to the model's own predicted crossing along each segment and ray, so the band scales with nucleus size and no dense label is needed. The second term is a duality loss adapted from RPCNet. RPCNet matches a boundary branch to the gradient of the predicted region mask, but trains that branch on contours taken from dense ground truth. Here the boundary head $e(x)$ has no ground truth. It is trained only to agree with $|\nabla \max_i m_i|$ and to stay sparse. This turns the masks' implied edges into a shared feature-level signal that sharpens all instances, and RPCNet's eight-iteration loop is dropped.
The design also addresses a tension in the literature. BASS argues that high-resolution features hallucinate jagged edges under noisy labels, and it downsamples to 14×14. Single-stage designs such as BlendMask and PointNu-Net argue that high resolution is needed to separate touching instances. Monotone ray and segment constraints are a smoothness prior that acts at full resolution. We test whether they let the model keep high-resolution features without the jaggedness BASS observed.
Baselines
The naive reference is a Voronoi-plus-color-clustering pseudo-mask, the label construction used by NINEPINS (Voronoi cells and k-means, then a modified HoVer-Net). We train our exact backbone and dynamic-filter head on these fixed pseudo-masks with the standard region loss. This isolates the effect of the label construction from the architecture, and our method must beat it on the touching-nucleus subset to justify itself.
The strongest click-supervised competitor is DoNuSeg. It refines pseudo-labels online from class-activation maps taken at the best encoder layer, adds activation-guided contrastive learning, and uses decoupled detection and segmentation heads. It reports AJI 3.9% above the next point-supervised method on CryoNuSeg, 2.9% on CoNSeP, and 2.2% on TNBC, but the file gives no absolute AJI values. No reported absolute result was found on Lacuna for DoNuSeg, so we will rerun it with its released settings and give it the same click sets, backbone family where possible, and an equal hyperparameter-search budget.
The strongest absolute numbers in the file come from the embedding method of Weakly Supervised Nuclei Segmentation via Instance Learning, which reports AJI 0.5430 on MultiOrgan and 0.5854 on TNBC. That method needs about 6 seconds of mean-shift clustering per image. Beating it means exceeding both AJI values under the same click protocol at under a tenth of its inference time.
As a color-edge competitor we reimplement the Boundary Self-Prediction module of P2MNet, which snaps masks to low-level color similarity and reaches 23.1 mask AP on COCO from points. We place it on our head in place of the handover loss.
Two references bound the problem from above. One is fully supervised PointNu-Net trained with full masks. The other is BASS with half the masks, which reports PQ 49.66% and Dice 76.76% on CryoNuSeg. Approaching BASS with clicks only would be a strong result, and exceeding it is not expected.
Ablations remove, one at a time, the handover term, the ray term, the adaptive contrastive band (replaced by BASS's fixed offset), and the duality head. The key ablation fixes the crossing at the segment midpoint. That is Voronoi expressed as a constraint, and it tests whether learning where the edge falls is what matters.
Why not just…?
Why not just make Voronoi pseudo-masks and train a standard single-stage model on them? This is the real threat. On isolated or well-separated nuclei it may be nearly as good. DoNuSeg reports that static Voronoi and level-set labels produce merged, blobby masks in low-contrast tissue. NINEPINS shows about a 6% drop in Dice and detection quality even when 30% of labels are true masks. Neither result isolates single-stage training on these labels. Our naive baseline and midpoint ablation do exactly that, with results reported separately for touching and isolated nuclei. If the gain appears only on isolated nuclei, or nowhere, the handover idea has failed and we will say so.
Why not just prompt SAM with the clicks? Work in the file indicates that off-the-shelf SAM does not resolve nuclei from weak prompts without dedicated adaptation, as described in All-in-SAM and BoxCell. It also needs a click per nucleus at test time, whereas our model needs clicks only for training. We include zero-shot point-prompted SAM, which is free to run, to quantify the gap.
Why not just refine coarse predictions with a DenseCRF? A DenseCRF snaps masks to color discontinuities. In this setting those discontinuities are weak or come from internal texture. Scribble-supervised saliency work needed the CRF inside a retraining loop with an edge network to be competitive, not as an inference-only fix. We add a DenseCRF to the naive baseline and report both accuracy and added latency.
Why not just use DoNuSeg? DoNuSeg's activation maps are spatially coarse and can spread into neighboring nuclei in crowded regions. It also has no explicit handover constraint between clicked neighbors. The head-to-head comparison stratified by contact settles whether this matters.
Proposed Experiments
Training clicks are simulated from ground-truth centroids with jitter up to the shift tolerance NINEPINS studied. Experiments run on CryoNuSeg, chosen for frozen-section low contrast; on CoNSeP and TNBC, chosen for clustering; and on MultiOrgan/MoNuSeg. PanNuke is used for multi-tissue generalization, with fully supervised FCIS at PQ 0.610 as a ceiling reference.
Metrics are AJI and panoptic quality, split into detection quality and segmentation quality, plus a boundary F-score at small pixel tolerances and images per second on one GPU. Every metric is reported separately for nuclei that touch a neighbor and for isolated nuclei, and for images grouped into terciles of local foreground–background intensity contrast.
Three outcomes would be decisive. First, the gain over the Voronoi-trained baseline should concentrate in the touching, low-contrast stratum, which is the regime the method targets. Second, the learned-crossing model should beat the midpoint ablation; otherwise learning the edge location adds nothing. Third, at full feature resolution the method should match or exceed a variant downsampled in the BASS manner, without the jagged edges BASS reports. A final check uses real rather than simulated clicks on a CryoNuSeg subset to confirm that the results survive human click noise.
References
- PointNu-Net: Keypoint-Assisted Convolutional Neural Network for Simultaneous Multi-Tissue Histology Nuclei Segmentation and Classification
- Boundary-aware Contrastive Learning for Semi-supervised Nuclei Instance Segmentation
- Joint Semantic Segmentation and Boundary Detection Using Iterative Pyramid Contexts
- BlendMask: Top-Down Meets Bottom-Up for Instance Segmentation
- NINEPINS: Nuclei Instance Segmentation with Point Annotations
- Dynamic Pseudo Label Optimization in Point-Supervised Nuclei Segmentation
- Weakly Supervised Nuclei Segmentation Via Instance Learning
- P2Object: Single Point Supervised Object Detection and Instance Segmentation
- All-in-SAM: from Weak Annotation to Pixel-wise Nuclei Segmentation with Prompt-based Finetuning
- BoxCell: Leveraging SAM for Cell Segmentation with Box Supervision
- Weakly-Supervised Salient Object Detection via Scribble Annotations
- The Four Color Theorem for Cell Instance Segmentation