End-to-End Cell Detection via Instance-aware Graph Modeling

• Bounding-box instance detection balances richer localization than point methods with lower annotation and computational overhead than segmentation.Source: §1 · p. 2
• Query-based detectors can struggle in dense or heterogeneous regions, and pathology-specific geometric methods require pixel-level annotations or decoupled training.Source: §1 · p. 2
• Prior two-stage graph methods require separate stage optimization, while feature-map graphs can obscure instance information with dense, noisy structures.Source: §1 · p. 2
• Learnable queries serve as candidate cell instances and graph nodes, with connectivity jointly determined by spatial proximity and feature similarity.Source: §1 · p. 2
• The query feature network maps patch-level context to instance-level semantics, while DGC builds the graph and IGN models inter-cell interactions.IGN reorganizes and filters queries before graph aggregation, then uses visual features to drive state updates along the cell topology.Source: §1 · p. 3
• The framework achieves state-of-the-art performance across multiple public benchmarks for cell and nucleus detection under different staining protocols.Source: §1 · p. 3

2.1. Instance Detection for Cells and Nuclei

• Bounding-box instance detection offers richer localization than point detection while avoiding the annotation and computational overhead of full segmentation.Source: §2.1 · p. 3

2.2. Graph-Based Learning in Pathology

• Two-stage graph construction derives nodes from detected or segmented cells, whereas end-to-end pixel- or patch-level graphs can contain redundant or biologically irrelevant edges.Source: §2.2 · pp. 3–4
• GCNs have limited receptive fields and can over-smooth, while standard GAT attention scales quadratically with node count; Mamba offers global modeling at linear complexity.Source: §2.2 · p. 4

3. Method

• The full framework maps patch features to instance-level features, builds an adaptive cell graph, and models local and global relational context alongside visual evidence.These components are designed to improve query-feature discrimination for instance separation and class prediction.Source: §3 · p. 4

3.1. Query Feature Learning Network

• Queries gather multi-scale visual and tissue context, are iteratively refined by a Transformer decoder, and provide spatial reference points for graph construction.Initial proposals are selected after class-agnostic NMS; between decoder layers, another NMS step reduces overlapping queries.Source: §3.1 · pp. 4–5

3.2. Dynamic Graph Construction

• DGC constructs query-node connections using a distance that combines spatial proximity and feature similarity, allowing graph structure to update with the queries.Spatial distance uses reference points, while feature distance uses L2-normalized query embeddings and cosine similarity.Source: §3.2 · p. 5

3.3. Instance-aware Graph Network

• IGN combines selective feature reorganization and graph aggregation with a topology-structured state-space layer driven by instance-level query features.CGSE uses all-query context to steer parameters in both SFR and TSS.Source: §3.3 · p. 6

3.3.1. Instance-aware Graph Learning

• SFR filters redundant queries and maps retained features into directional representations; graph aggregation combines them into a topological representation where dissimilar neighbors can remain non-interfering.The method uses the geometry of representations for selectivity rather than explicit edge pruning.Source: §3.3.1 · pp. 6–7

3.3.2. Topology-structured State Space

• TSS uses graph-aggregated topology as the incoming state and repeatedly injects visual features, combining relational and appearance evidence in a single state.The resulting features feed classification and bounding-box regression heads; predictions are matched to ground truth with the Hungarian algorithm.Source: §3.3.2 · pp. 7–8

3.3.3. Context-Guided State Encoding.

• CGSE generates state-space parameters from instance features, with a globally shared A initialized using HiPPO-LegS and B and C learned from cellular knowledge.A low-rank structure is used for Δ to control parameter count.Source: §3.3.3 · p. 8

4.1. Datasets

• Evaluation spans CoNSeP, CytoDArk0, and OCELOT, covering colorectal adenocarcinoma, mammalian brain tissue, and cell nuclei from TCGA whole-slide images.The datasets include H&E- and Nissl-stained images with nucleus-level or cell-level annotations.Source: §4.1 · p. 8
• Patch sizes are 128×128 for denser CoNSeP images and 256×256 for CytoDArk0 and OCELOT.Annotations were converted to bounding-box labels for evaluation.Source: §4.1 · p. 8

4.2. Implementation Details

• Experiments use ResNet-50, and general DETR-based methods share 900 queries for fair comparison.The reported setup also uses an NVIDIA A40 GPU and AdamW with an initial learning rate of 2e-4 and weight decay of 0.05.Source: §4.2 · p. 8

4.3. Evaluation Setup and Metrics

• Detection-method comparisons use mAP, mAP50, mAP75, and AR across confidence thresholds.These metrics assess detection and classification performance for detection methods.Source: §4.3 · p. 9
• Cell-specific classification methods are compared using F1-score because their supervision and output granularity differ from detection methods.The paper follows in using F1-score as a consistent classification measure.Source: §4.3 · p. 9
• The comparison includes general-purpose and pathology-specific detectors, plus classification methods on the multi-class dataset.The classification comparison is intended to assess performance against pathology-specific models.Source: §4.3 · p. 9

4.4.1. Comparison with Detection Methods

• On CoNSeP, the method surpasses DDQ-DETR by 2.0% in mAP and 2.2% in AR; gains over DDQ-DETR are 2.1% in mAP and 1.4% in AR on CytoDArk0.The authors report superiority across all metrics on OCELOT and statistically significant improvements (p < 0.05).Source: §4.4.1 · p. 10
• Across Nissl- and H&E-stained examples, the model detects more cells and reduces false positives, particularly for densely packed and adherent cells.Figure 4 compares DINO, CO-DINO, DDQ-DETR, and the proposed model; panels A–B are Nissl-stained and C–E are H&E-stained.Source: §4.4.1 · p. 10, Figure 4 · p. 10

4.4.2. Comparison with Classification Methods

• The method achieves state-of-the-art performance across reported classification metrics, outperforming mask-based and point-based approaches, including on the challenging rare class.It attains the highest F_cm, which varies substantially across methods; the authors note the rare class’s scarcity and atypical morphology.Source: §4.4.2 · pp. 10–11, Table 2 · p. 11

Ablation study

• Adding IGL to the baseline raises mAP by 0.8%, and adding TSS yields a further 1.2% gain; IGN also outperforms the tested alternative graph learners.The baseline is 30.0% mAP and 56.5% AR.Source: §4.5 · p. 11
• Removing positional similarity reduces mAP from 32.0% to 31.4% and AR from 58.7% to 57.3%.Source: §4.5 · p. 11
• Removing feature similarity lowers mAP to 30.6% and AR to 56.6%, larger drops than removing positional similarity.The reported decreases are 1.4% in mAP and 2.1% in AR.Source: §4.5 · p. 11

Efficiency analysis

• The model uses 49.57M parameters and 276 GFLOPs, and its graph refinement adds only 1.25M parameters and 6 GFLOPs over DDQ-DETR.The authors describe it as substantially lighter than instance-segmentation methods and comparable to point-based models, while providing a more complex task than point detection.Source: §4.6 · pp. 11–12, Table 4 · p. 12

4.7.1. Investigation of 𝛼-values

• The query-distance weighting achieves optimal performance at α = 0.5, with spatial and feature distance contributing equally and complementarily.The model adopts α = 0.5.Source: §4.7.1 · p. 12

4.7.2. Investigation of Query Numbers

• Increasing queries from 100 to 300 raises AR by 2.3% and mAP50 by 1.9%; increasing to 900 adds 3.8% and 2.3%, respectively.Larger query counts continue to improve AR but decrease precision metrics, so the authors adopt 900 queries; larger values were not tested due to GPU memory constraints.Source: §4.7.2 · p. 12

4.7.3. Investigation of K-values

• The preferred K varies by dataset: CoNSeP performs best at K = 8, CytoDArk0 at K = 3, and OCELOT at K = 4.On densely packed, highly overlapping CoNSeP nuclei, K = 8 outperforms sparser K = 5 by up to 2.5% in mAP; the study tests sparse and dense connectivity regimes.Source: §4.7.3 · p. 13, Table 5 · p. 13

5. Discussion

• Cells are shaped by interactions in their microenvironment, while appearance-focused detectors can falter in dense, overlapping, or heterogeneous regions and post-hoc graphs inherit upstream errors.The discussion motivates making relational reasoning intrinsic to detection rather than appending it afterward.Source: §5 · p. 13
• Learnable queries serve as candidate-cell graph nodes, carrying decoded multi-scale visual evidence from patch representations into instance-level relational reasoning.This single representation lets feature extraction, instantiation, and graph reasoning be optimized under a unified objective.Source: §5 · p. 13
• Graph connectivity combines spatial proximity and feature similarity, with query updates and adjacency shaping each other during learning.The resulting topology adapts to data rather than remaining fixed by spatial heuristics.Source: §5 · p. 14
• A selective state-space model suppresses irrelevant signals in redundant query graphs while preserving global structure at linear node complexity, unlike dense attention’s quadratic cost.Its topological latent state is driven by re-injected query visuals, and CGSE generates transition parameters from instance-level context.Source: §5 · p. 14
• NMS and KNN act as parameter-free structural selectors; learned query embeddings, rather than changes within KNN, make graph connectivity dynamic.Retained query features still participate in backpropagation, so these selectors do not interrupt end-to-end training.Source: §5 · p. 14

6. Conclusion

• Learnable queries act as graph nodes, and spatial and feature cues dynamically construct a context-aware topology for varying cell distributions.The framework is designed to improve instance discrimination in query-based detection through structured reasoning.Source: §6 · p. 14
• The instance-aware graph network uses a Topology-structured State Space model with graph topology as its latent state to address noise and long-range dependencies in large query sets.Source: §6 · p. 14
• The method achieves state-of-the-art performance across multiple benchmarks.Source: §6 · p. 14
• The paper presents combining query-based detection with structured graph learning as a promising direction for precise and robust cell instance analysis in digital pathology.Source: §6 · p. 14