Research map / Detection and segmentation / CNN features
DPA-P2PNet: Deformable Proposal-Aware P2PNet for Accurate Point-Based Cell Detection
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes DPA-P2PNet, an end-to-end point-based cell detector that decodes multi-scale features at point proposal coordinates and learns deformable proposals, plus an mFoV variant and the first SSL pre-training on a large IHC pathology dataset.
Problem
Point-based cell detection in histopathology: P2PNet decodes only from a single feature level with scale-agnostic point proposals that are spatially biased from cell centroids, and cannot exploit multi-field-of-view context for cell classification.
Approach
Extract ROI features from each point proposal's coordinates across the feature pyramid (making proposals proposal-aware), learn implicit offsets from the finest level to deform proposals toward cells, and fuse larger-FoV pyramid features into the small-FoV pyramid for mFoV decoding.
Key results
- On CoNSeP, DPA-P2PNet reaches F1 71.1 and AP 62.9 versus P2PNet's 70.0 and 60.5.
- On OCELOT, mFoV DPA-P2PNet with K=2 (transposed convolution) reaches F1 62.6 and AP 49.1, surpassing MFoVCE-Net+ by 2.8% F1 and 4.9% AP.
- With ViT-B/16, MAE pre-training on the IHC PD-L1 data reaches F1 61.7 and AP 51.6, beating ImageNet-supervised init by 3.7% F1 and 4.6% AP.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add multi-scale decoding (MSD) (ablation) | P2PNet · CoNSeP | F1: 70.0 -> 70.7 | Helped |
| Add multi-scale decoding (MSD) (ablation) | P2PNet · CoNSeP | AP: 60.5 -> 61.7 | Helped |
| Add deformable point proposals (DPP) to MSD (ablation) | DPA-P2PNet with MSD · CoNSeP | F1: 70.7 -> 71.1 | Helped |
| Add deformable point proposals (DPP) to MSD (ablation) | DPA-P2PNet with MSD · CoNSeP | AP: 61.7 -> 62.9 | Helped |
| Use 2 FoVs instead of 1 (ablation) | mFoV DPA-P2PNet · OCELOT | F1: 59.3 -> 62.6 | Helped |
| Use 2 FoVs instead of 1 (ablation) | mFoV DPA-P2PNet · OCELOT | AP: 44.4 -> 49.1 | Helped |
| Increase the input from 1 to 4 FoVs (ablation) | mFoV DPA-P2PNet · OCELOT | F1: 59.3 -> 64.6 | Helped |
| Increase the input from 1 to 4 FoVs (ablation) | mFoV DPA-P2PNet · OCELOT | FPS: 25 -> 8 | Hurt |
| Use MoCo v2 instead of ImageNet-supervised initialization (ablation) | DPA-P2PNet with ResNet-50 · PD-L1 | F1: 55.9 -> 57.4 | Helped |
| Use MAE instead of ImageNet-supervised initialization (ablation) | DPA-P2PNet with ViT-B/16 · PD-L1 | F1: 58.0 -> 61.7 | Helped |
Surprising findings
- ImageNet supervised pre-training is outperformed by domain-specific SSL on IHC pathology data for both ResNet-50 and ViT-B/16 backbones.
- Increasing the number of mFoV inputs improves accuracy but reduces inference speed by 68% at K=4.
Limitations
- Gains over P2PNet on some datasets are modest (e.g., 1.5% F1 on PD-L1) and the added cost in parameters and MACs is non-trivial.
- mFoV inference speed drops substantially as the number of FoVs grows, with 8 FPS at K=4.
- Evaluation uses an internal PD-L1 dataset that is not publicly available.
Experimental setup
Datasets: CoNSeP BCData PD-L1 (internal IHC) OCELOT IHC pre-training data (Ki-67, PD-L1, HER2) ImageNet-1K
Benchmarks: CoNSeP BCData PD-L1 OCELOT
Metrics: F1-score AP Params (M) MACs (G) FPS
Baselines: U-Net DeepLabV3+ U-CSRNet MCSpatNet ML-CellNet GL FIDT OT-M P2PNet CLTR E2E MFoVCE-Net
Builds on: P2PNet FPN ViT-Adapter MoCo v2 DINO
Closest papers by meaning
- M2OST: Many-to-one Regression for Predicting Spatial Transcriptomics from Digital Pathology Images (AAAI 2025 · similarity 0.82)
- MEATRD: Multimodal Anomalous Tissue Region Detection Enhanced with Spatial Transcriptomics (AAAI 2025 · similarity 0.82)
- PathVQ: Reforming Computational Pathology Foundation Model for Whole Slide Image Analysis via Vector Quantization (NeurIPS 2025 · similarity 0.82)
- Towards Scalable Context-Aware Single-Cell Spatial Transcriptomics Prediction from Histology Images (NeurIPS 2026 · similarity 0.81)
- GeneFlow: Translation of Single-cell Gene Expression to Histopathological Images via Rectified Flow (NeurIPS 2025 · similarity 0.81)
- Cell Graph Transformer for Nuclei Classification (AAAI 2024 · similarity 0.81)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.