HistoPLUS: Advanced Computational Pathology

• HistoPLUS is a state-of-the-art computational framework for comprehensive cellular analysis in H&E stained tissues, excelling in cell detection, segmentation, and classification.
• It builds on a hybrid architecture combining a compact H0-mini transformer with a ViT encoder and HoVerNet-style decoder, enhanced by advanced annotation and post-processing techniques.
• Quantitatively validated on large pancancer datasets, HistoPLUS demonstrates superior detection quality, robust zero-shot domain transfer, and efficient integration in digital pathology pipelines.

HistoPLUS is a state-of-the-art computational framework for comprehensive cellular analysis in hematoxylin and eosin (H&E) histopathology, specifically designed for cell detection, segmentation, and classification across diverse tissue types, including tumor microenvironments and inflammatory lesions. Developed and quantitatively validated on large, expertly curated pancancer datasets, it achieves robust performance on rare and understudied cell phenotypes, while remaining efficient in both computational footprint and inference speed. HistoPLUS forms the core cellular characterization module in advanced digital pathology pipelines and has demonstrated strong cross-domain transfer, supporting applications from tumor microenvironment mapping to interpretable multiple instance learning for inflammation biomarker extraction (Adjadj et al., 13 Aug 2025, Baiocco-Rodrigues et al., 15 Dec 2025).

1. Model Architecture and Algorithmic Innovations

HistoPLUS builds upon the CellVIT panoptic instance segmentation architecture with a ViT encoder and a task-multiplexed HoVerNet-style decoder (Adjadj et al., 13 Aug 2025). Its backbone is the “H0-mini” transformer (≈86M parameters), distilled from a large pathology foundation model (H-Optimus-0) via self-supervised DINO and iBOT objectives, yielding a pathology-specific embedding space with high morphologic and staining fidelity. The framework comprises three heads:

• Nuclei Prediction (NP): Binary boundary segmentation for cell delineation.
• Horizontal-Vertical (HV): Estimation of normalized horizontal/vertical distances to each nuclear centroid, facilitating spatial separation.
• Nuclei Type (NT): Per-instance classification into 13 morphologically and functionally distinct cell types (e.g., epithelial, lymphocyte, fibroblast, neutrophil, etc.).

Fused multi-scale features from four encoder depths are routed to each decoder head, enabling fine boundary localization and robust phenotype assignment. HoVerNet-derived post-processing, using Sobel gradient filters and marker-controlled watershed, yields sharply separated, instance-resolved nuclei, minimizing merged or fragmented detection errors. The resulting architecture is five times smaller than standard ViT-Huge backbones employed in similar pipelines (e.g., SAM-H, UNI2).

2. Curation and Annotation Strategy for Training Data

HistoPLUS was trained on HistoTRAIN, a curated dataset of 108,722 manually annotated nuclei from 739 H&E-stained whole-slide images spanning six major cancer types (bladder, colon, lung adenocarcinoma, lung squamous cell carcinoma, mesothelioma, and pancreatic) (Adjadj et al., 13 Aug 2025). The annotation pipeline integrates:

1. Active learning-driven tile selection:
   • K-means clustering on Phikon features for coverage of morphological diversity.
   • Shallow MLP classifiers trained to predict rare cell types for undersampled phenotype selection (eosinophils, neutrophils, mitoses).
   • BALD uncertainty sampling to prioritize ambiguous regions likely to contain hard-to-classify events.
2. Expert point annotations:
   • Pathologists mark centroids and assign one of 13 class labels using Cytomine.
3. Segmentation refinement:
   • NuClick converts point labels to precise nuclear masks.

Cell type representation was actively balanced, resulting in improved detection for rare classes (e.g., 1.68% eosinophils, 5.88% neutrophils). This curation enables high-fidelity classification and detection generalization across multiple tissue and indication domains.

3. Training Objectives and Optimization

The HistoPLUS training objective is a weighted sum of three branch-specific losses applied patch-wise:

$$L_\text{total} = w_\mathrm{NP} L_\mathrm{NP} + w_\mathrm{HV} L_\mathrm{HV} + w_\mathrm{NT} L_\mathrm{NT}$$

• Nuclei Prediction (NP): Combined binary cross-entropy and Dice loss for segmentation accuracy.
• HV loss: Mean-squared error with gradient consistency penalty for spatial localization.
• Nuclei Type (NT): Focal Tversky loss, which controls false positive/negative penalties and mitigates class imbalance.

Weights $(w_\mathrm{NP},w_\mathrm{HV},w_\mathrm{NT})$ follow empirically optimized values determined in the original CellVIT study. This composite loss enforces consistent nuclear boundary detection, centroid localization, and phenotype assignment.

4. Quantitative Performance and Cross-Domain Validation

External validation demonstrates that HistoPLUS achieves or surpasses prior state-of-the-art performance across benchmarks (Adjadj et al., 13 Aug 2025, Baiocco-Rodrigues et al., 15 Dec 2025):

Metric	HistoPLUS (H0-mini)	SOTA Baseline (SAM-H)	Δ (HistoPLUS – SOTA)
Detection Quality (DQ)	0.725	0.688	+0.052
Segmentation Quality (SQ)	0.801	0.808	–0.007
Average Classification F1 (13 types)	0.568	—	+23.7% (relative)

Improvements are statistically significant (p < 1e-3) for 8 of 13 cell types, with marked gains in epithelial (F1 0.422 vs. 0.223), smooth muscle, mitotic figure, and endothelial cell classes.

Zero-shot transfer to unseen cancer indications (ovarian, breast) yields:

• Breast: DQ = 0.836, SQ = 0.801; F1 = 0.799 (lymphocyte), 0.475 (cancer cell)
• Ovarian: DQ = 0.805, SQ = 0.803; F1 = 0.682 (cancer cell), 0.639 (lymphocyte)

This confirms the domain-agnostic robustness and practical value for large-scale digital pathology biomarker extraction.

5. Integration in Downstream Analytical Pipelines

HistoPLUS is integrated in IMILIA, an interpretable multiple instance learning framework for inflammation prediction in IBD (Baiocco-Rodrigues et al., 15 Dec 2025). The workflow:

• Top-scoring H&E tiles (224×224 px) are selected via MIL (Chowder).
• Each tile is processed by HistoPLUS for panoptic cell instance segmentation and 13-way classification.
• EpiSeg delineates epithelial regions; cell densities of immune subtypes within the epithelium are calculated:

$$\rho^c = \frac{\sum_{k=1}^{N^c} \mathbf{1}\left(E(c_{k,x},c_{k,y})=1\right)}{(\sum_{x,y} E(x,y)) \cdot mpp_x \cdot mpp_y}$$

where $E(x,y)$ is the binary epithelium mask and $mpp_x, mpp_y$ are microns-per-pixel resolutions.

Per-class F1 and panoptic detection/segmentation metrics are reported for both oncology and IBD domains. For instance, in IBD transfer (SPARC IBD), epithelial F1 reaches 0.70, lymphocyte 0.53; detection quality 0.774, segmentation quality 0.755. A strong Pearson correlation ($r=0.85, p<10^{-8}$) exists between HistoPLUS epithelial counts and EpiSeg area, underscoring measurement consistency.

Notably, HistoPLUS supports biomarker discovery by quantifying neutrophil and lymphocyte infiltration in distinct tissue compartments without retraining, utilizing the pre-trained multiclass pan-cancer model.

6. Practical Deployment, Code Availability, and Recommendations

Open-source inference scripts and model weights are provided at https://github.com/owkin/histoplus under CC-BY-4.0 (Adjadj et al., 13 Aug 2025). The recommended deployment procedure:

• Tile WSIs at 40× (448×448 patches).
• Apply CellVIT-HistoPLUS (FP16 compatible).
• Perform HoVerNet post-processing for instance extraction.
• Aggregate outputs for downstream statistical analysis.

A single Tesla T4 GPU processes a 1 cm² region in 20–40 minutes. Macenko color normalization is recommended for staining consistency; batch inference with DeepSpeed or AMP is supported. All post-processing dependencies (NuClick, Cytomine, foundation models) retain their original licenses.

The modular architecture allows for pipeline integration in C#/.NET or Python, with output formats including 2D label masks and cross-sectional cell tables.

7. Comparative Impact and Significance

HistoPLUS advances the state of the art in digital pathology, primarily by enabling accurate, label-efficient detection and classification of rare cell types, facilitating applications in tumor microenvironment research and translational inflammation pathology. Empirical results substantiate its superiority in detection quality and class-balanced F1 over previously dominant models, while requiring significantly fewer parameters and resources. Its demonstrated zero-shot domain transfer supports broad applicability for quantitative histopathology studies, forming the basis for interpretable and reproducible cellular phenotype quantification in clinical and research pipelines (Adjadj et al., 13 Aug 2025, Baiocco-Rodrigues et al., 15 Dec 2025).