> For the complete documentation index, see [llms.txt](https://www.parapathology.com/llms.txt). Markdown versions of documentation pages are available by appending `.md` to page URLs; this page is available as [Markdown](https://www.parapathology.com/clippings/histoplus-towards-comprehensive-cellular-characterisation-of-h-and-e-slides.md). # HistoPLUS: Towards Comprehensive Cellular Characterisation of H\&E Slides HistoPLUS combines an active-learning pan-cancer dataset (HistoTRAIN: 108k nuclei, 13 cell types, 6 indications) with a compact CellViT architecture powered by a distilled pathology foundation model ( ## Summary Accurate single-cell detection, boundary segmentation, and phenotypic classification on standard hematoxylin and eosin (H\&E) whole-slide images (WSIs) are prerequisites for quantifying the tumor microenvironment (TME). Profiling spatial cellular interactions—such as tumor-associated neutrophil (TAN) infiltration, cancer-associated fibroblast (CAF) stromal remodeling, plasma cell clustering, and microvascular invasion—directly predicts disease progression and immunotherapy responsiveness. However, existing computational pathology methods face two crippling bottlenecks: 1. **The Understudied Cell-Type Deficit in Public Corpora:** Landmark datasets (PanNuke, CoNSeP, Lizard, MoNuSAC) either provide segmentation masks without lineage labels, provide only coarse bounding boxes, or restrict classifications to 4–6 broad categories (neoplastic, healthy epithelial, inflammatory, connective/stroma, dead). Clinically critical non-lymphoid immune populations (neutrophils, eosinophils, macrophages), plasmocytes, vascular/pericytic elements (endothelium, smooth muscle), and key cytologic events (mitotic figures, apoptotic bodies) remain absent or heavily underrepresented. 2. **The Foundation Model Scale Paradox:** Pretrained pathology foundation models (PFMs) such as UNI2 (681M) and Virchow2 (632M) or vision models like SAM-Huge (636M) incur massive compute, memory, and latency penalties. Yet for cellular instance segmentation, parameter scaling exhibits steep diminishing returns: segmentation quality improvements hit a plateau ($\Delta SQ \le 1.5$), leaving smaller clinics and high-throughput research pipelines bottlenecked by GPU compute. Published in the *Journal of Pathology Informatics* (Volume 22, 2026, 100696, PII [S2153-3539(26)00156-2](https://www.sciencedirect.com/science/article/pii/S2153353926001562), [DOI: 10.1016/j.jpi.2026.100696](https://doi.org/10.1016/j.jpi.2026.100696); [arXiv:2508.09926](https://arxiv.org/abs/2508.09926); local PDF: [1-s2.0-S2153353926001562-main.pdf](file:///K:/DownloadsK/1-s2.0-S2153353926001562-main.pdf)) by Benjamin Adjadj, Pierre-Antoine Bannier, Guillaume Horent, Sebastien Mandel, Gary Klajer, Aurore Lyon, Kathryn Schutte, Ulysse Marteau, Valentin Gaury, Laura Dumont, Thomas Mathieu, Reda Belbahri, Benoît Schmauch, Eric Durand, Katharina Von Loga, and Lucie Gillet across **Owkin France**, **Bioptimus**, and the **MOSAIC consortium**, **HistoPLUS** resolves both dilemmas through an active-learning curation engine and a distilled foundation-model architecture: * **HistoTRAIN & Active-Learning Curation (108,722 Nuclei):** Curated across 739 WSIs spanning 6 cancer indications (bladder, colon, lung adenocarcinoma, lung squamous cell carcinoma, mesothelioma, and pancreatic adenocarcinoma). Rather than passive labeling, an active-learning pipeline deployed 3 parallel arms (Phikon feature diversity clustering, lightweight rare-cell MLP detectors, and BALD epistemic uncertainty sampling) combined with \[\[NuClick]] automated boundary propagation from expert point annotations. * **HistoVAL Multi-Reader Consensus Benchmark (69,108 Nuclei):** Established across 530 ROIs ($112 \times 112\ \mu\text{m}$ at $40\times$) from 248 slides across 6 independent cohorts from the MOSAIC consortium. Ground truth was forged via Hungarian instance matching (IoU $>0.4$) across 3 independent pathologist annotations per tile, retaining only consensus nuclei ($\ge 2$ agreeing pathologists) with majority-voted phenotypes. Includes 2 completely unseen cancer indications: ovarian serous cystadenocarcinoma (OV) and breast invasive carcinoma (BRCA). * **Distilled PFM Integration within CellViT (H0-mini):** Integrates the compact **H0-mini** foundation model (86M parameters, distilled from Bioptimus' H-Optimus-0 via joint DINO class-token and iBOT masked patch distillation) as the ViT encoder inside a 3-branch CellViT decoder (nuclear prediction, horizontal-vertical distance regression, and 14/15-class nuclei typing with Focal Tversky loss). * **Empirical Superiority:** Outperforms competing state-of-the-art models by **+5.2% in Detection Quality (DQ)** and **+23.7% in overall classification F1 score**, achieving statistically significant gains on 8 of 13 cell types and unlocking 7 understudied lineages. Matches the panoptic and detection quality of ViT-Huge models (UNI2, Virchow2) while utilizing **$5\times$ fewer parameters** and running **$2.1\times$ faster**. * **Zero-Shot Indication Generalization:** Robustly transfers to unseen breast cancer (DQ 0.836, SQ 0.801, Lymphocyte F1 0.799) and ovarian cancer (DQ 0.805, SQ 0.803, Cancer Cell F1 0.682). * **Production Open Source:** Complete inference CLI, Python API, and GeoJSON QuPath export pipelines are released on GitHub ([owkin/histoplus](https://github.com/owkin/histoplus)), with checkpoints available on Hugging Face ([Owkin-Bioptimus/histoplus](https://huggingface.co/Owkin-Bioptimus/histoplus)). *** ## Architectural Blueprint & Inference Mechanics ```typescript ┌────────────────────────────────────────────────────────────────────────────────────────────────────────┐ │ HISTOPLUS INFERENCE PIPELINE │ └────────────────────────────────────────────────────────────────────────────────────────────────────────┘ [ Whole Slide Image (H&E: .svs, .ndpi, .tiff) ] │ ▼ ┌──────────────────────────────────────────────────────────────────────────────────────────────────────┐ │ 1. DEEPZOOM TILING & PRE-PROCESSING │ │ • MPP Selection: 0.25 µm/px (40×, tile size 448 px) or 0.50 µm/px (20×, tile size 224 px) │ │ • Inference Grid: 784 × 784 px tiles (multiples of ViT patch sizes 14 & 16) │ │ • Border Overlap: 64 px overlap between consecutive tiles to eliminate edge boundary artifacts │ │ • Normalization: Bioptimus Mean (0.707, 0.579, 0.704), Std (0.212, 0.230, 0.178) │ └──────────────────────────────────────────────────┬───────────────────────────────────────────────────┘ │ ▼ ┌──────────────────────────────────────────────────────────────────────────────────────────────────────┐ │ 2. DISTILLED PFM ENCODER (Bioptimus H0-mini ViT-Base Backbone — 86M Params) │ │ • Distilled from H-Optimus-0 via DINO global class token + iBOT local masked patch supervision │ │ • Dynamic 2D positional embedding interpolation accommodating multi-scale inputs │ │ • Multi-Scale Feature Extraction: Skips tapped from Transformer layers 3, 5, 7, and 11 │ └──────────────────────────────────────────────────┬───────────────────────────────────────────────────┘ │ Multi-scale feature tokens ▼ ┌──────────────────────────────────────────────────────────────────────────────────────────────────────┐ │ 3. MULTI-TASK CELLVIT DECODER (Isolated Upsampling Branches) │ │ ├── [Branch 1: Nuclei Prediction (NP)] ──► Binary boundary mask (Weighted BCE + Dice Loss) │ │ ├── [Branch 2: Horizontal-Vertical (HV)] ─► Distance gradients (x, y) to centroid (MSE Loss) │ │ └── [Branch 3: Nuclei Type (NT)] ─────────► 15-class instance logits (Focal Tversky Loss) │ └──────────────────────────────────────────────────┬───────────────────────────────────────────────────┘ │ ▼ ┌──────────────────────────────────────────────────────────────────────────────────────────────────────┐ │ 4. TWO-STAGE HOVERNET WATERSHED POST-PROCESSING │ │ • Sobel filtering on horizontal & vertical normalized gradient vectors │ │ • Nuclear mountain inversion (distance basin mapping with Gaussian blur 3×3) │ │ • Marker-controlled watershed instance boundary separation │ │ • Majority voting on Nuclei Type (NT) channel within each segmented instance mask │ └──────────────────────────────────────────────────┬───────────────────────────────────────────────────┘ │ ▼ ┌────────────────────────────────────────────────────────┐ │ Outputs: Polygon Contours, Centroids, Bounding Boxes │ │ + 13-Class Phenotypic Labels & Prediction Confidence │ │ (Exportable to Native JSON and QuPath-Ready GeoJSON) │ └────────────────────────────────────────────────────────┘ ``` ### Why H0-mini Defeats ViT-Huge Foundation Models for Single-Cell Tasks A central finding of the HistoPLUS evaluation is that scaling vision transformer parameters from 86M (ViT-Base) to $>600\text{M}$ (ViT-Huge/Giant) yields negligible benefit for nuclear segmentation and classification: 1. **Segmentation Quality (SQ) Invariance:** Across all tested backbones (SAM-B, SAM-H, Phikon, Hibou-B, H0-mini, UNI2, Virchow2), the average segmentation quality (mean IoU of matched nuclei) varied by less than **$\Delta SQ \le 1.5%$**. Nuclear boundary delineation on H\&E is governed primarily by local gradient contrast (hematoxylin-rich chromatin vs. cytoplasmic eosin) captured effectively by high-resolution multi-scale skip decoders, rendering giant receptive fields redundant. 2. **Detection Quality Parity:** In external multi-cohort validation on HistoVAL ($n=530$), H0-mini matched both UNI2 and Virchow2 with an identical Detection Quality score of **0.753 \[0.742–0.763]**. 3. **The Compute & Latency Dividend:** By deploying H0-mini instead of UNI2 or Virchow2, HistoPLUS slashes parameter count by **$5.1\times$** and inference runtime by **$2.1\times$**. An entire surgical resection whole-slide image can be parsed in 20 to 40 minutes on a standard single NVIDIA Tesla T4 GPU, making whole-cohort TME analysis computationally accessible. *** ## Dataset Curation & Multi-Reader Ground Truth ```typescript ┌──────────────────────────────────────────────────────────────────────────────────────────────────────┐ │ HISTOPLUS DATA ENGINE & BENCHMARKS │ └──────────────────────────────────────────────────────────────────────────────────────────────────────┘ [ 739 TCGA Slides (6 Cancers) ] [ 248 MOSAIC Slides (6 Cohorts) ] │ │ ▼ ▼ ┌───────────────────────────┐ ┌───────────────────────────┐ │ 3-Arm Active Learning: │ │ Multi-Pathologist Review: │ │ 1. Phikon K-means (K=20) │ │ • 3 pathologists / tile │ │ 2. Rare-Cell MLPs │ │ • Point annotations │ │ 3. BALD Uncertainty (>90%)│ │ • NuClick boundary exp. │ └─────────────┬─────────────┘ └─────────────┬─────────────┘ │ │ ▼ ▼ ┌───────────────────────────┐ ┌───────────────────────────┐ │ Expert Point Annotations │ │ Hungarian Instance Match: │ │ on Cytomine Platform │ │ • IoU > 0.4 consensus │ │ │ │ │ • Retain >= 2 raters │ │ ▼ │ │ • Centroid average │ │ NuClick Contour Expansion │ │ • Majority class voting │ └─────────────┬─────────────┘ └─────────────┬─────────────┘ │ │ ▼ ▼ ┌───────────────────────────┐ ┌───────────────────────────┐ │ HistoTRAIN Training Set │ │ HistoVAL Validation Set │ │ • 108,722 nuclei │ │ • 69,108 consensus nuclei │ │ • 1,415 tiles (448 × 448) │ │ • 530 ROIs (112 × 112 µm) │ │ • 6 Training Indications │ │ • 4 Seen + 2 Unseen Cancers│ │ (BLCA, COAD, LUAD, │ │ (Adds Breast & Ovarian) │ │ LUSC, MESO, PAAD) │ └───────────────────────────┘ └───────────────────────────┘ ``` ### NuClick Point-Prompted Boundary Expansion & Multi-Reader Consensus Handmade polygon tracings are notoriously time-consuming (taking 30–60 seconds per nucleus) and exhibit marked inter-pathologist boundary discordance. To standardize boundary geometry and accelerate large-scale dataset generation, HistoPLUS deployed the deep learning framework \[\[NuClick]] ([mostafajahanifar/nuclick\_torch](https://github.com/mostafajahanifar/nuclick_torch/)): 1. **NuClick U-Net Pretraining:** * Deployed the vanilla NuClick U-Net architecture with encoder-decoder skip connections. * Pretrained on a combined corpus of **ConSEP** (colorectal adenocarcinoma) and **PCNS** (pan-cancer nuclei segmentation, spanning 14 distinct cancer indications). * Training patches were cropped around target nuclei ($30 \times 30$ to $45 \times 45$ px), conditioned on a 5-channel tensor containing RGB histology, a Gaussian target inclusion map, and neighbouring cell exclusion maps. * Trained with weighted Binary Cross-Entropy and Dice loss over 35 epochs, achieving a top validation **Dice score of 0.874**. 2. **The 4-Step Multi-Pathologist Consensus Protocol (HistoVAL):** * *Step 1 (Independent Point Clicks):* 3 out of 9 board-certified pathologists independently annotated centroid coordinates and attributed cell lineages on Cytomine across 530 candidate ROIs, collecting an initial pool of **212,992 raw point annotations**. * *Step 2 (Prompted Boundary Expansion):* The pretrained NuClick model inferred exact instance masks from each pathologist's point coordinates. * *Step 3 (Hungarian Instance Matching):* Nuclear masks between raters were paired using the Hungarian algorithm based on an Intersection over Union cutoff ($\text{IoU} > 0.40$). Only nuclei verified by a consensus of **$\ge 2$ agreeing pathologists** were retained; singleton or highly discordant annotations were discarded. * *Step 4 (Consensus Centroid Averaging & Final Masking):* For each consensus cluster, centroid coordinates were averaged across agreeing raters, and the phenotypic class was determined via majority voting. A final inference pass of NuClick was executed on the averaged consensus centroids, establishing the definitive ground truth of **69,108 consensus nuclei** across 530 ROIs. *** ### The 13 Cell-Type Taxonomy HistoPLUS expands cell characterization from broad 5-class categories to **13 clinically relevant cell lineages and morphological events**: | Broad Compartment | Granular Cell Type | Morphological & Biological Significance | | ----------------------------- | ---------------------------------- | --------------------------------------------------------------------------------------------------- | | **Neoplastic** | **Cancer cell** | Pleomorphic, enlarged nuclei, hyperchromasia, prominent nucleoli, high N:C ratio | | **Immune (Lymphoid)** | **Lymphocytes** | Small, round, hyperchromatic spherical nuclei, scant cytoplasm; sTILs titration | | | **Plasmocytes (Plasma cells)** | Eccentric clock-face chromatin, cartwheel pattern, perinuclear hof; chronic B-cell immunity | | **Immune (Myeloid / Innate)** | **Macrophages** | Large indented/bean-shaped nuclei, delicate foamy/vacuolated cytoplasm; phagocytic stroma | | | **Neutrophils** | Multi-lobed (3–5 lobes) polymorphonuclear morphology; acute inflammation, immunosuppressive TANs | | | **Eosinophils** | Bilobed nuclei, intense bright eosinophilic cytoplasmic granules; allergic/anti-parasitic reactions | | **Stroma & Vasculature** | **Fibroblasts** | Elongated, slender, spindle-shaped nuclei embedded within collagenous ECM; CAFs | | | **Endothelial cells** | Flattened, attenuated nuclei lining vascular lumen channels; angiogenesis profiling | | | **Muscle cells (Smooth muscle)** | Elongated cigar-shaped/blunt-ended nuclei organized in parallel bundles | | **Normal Tissue Context** | **Epithelial (Non-cancerous)** | Uniform, polarized cuboidal or columnar nuclei lining benign glandular or ductal structures | | | **Red blood cells (Erythrocytes)** | Anucleate biconcave discs, intense eosinophilic cytoplasm; vascular landmarks | | **Cytologic Dynamics** | **Mitotic figures** | Condensed, dark, hairy chromatin aggregates lacking nuclear envelope; proliferative index | | | **Apoptotic bodies** | Pyknotic, fragmented, condensed round chromatin spheres with halo; cellular turnover | *(Note: In implementation, an additional background class and minor stromal class are indexed, yielding 15 internal classes).* *** ## Benchmark Results & Empirical Findings ### 1. External Multi-Cohort Validation on HistoVAL ($n=530$) In external testing against the general-purpose vision foundation model (SAM) and leading pathology models across 4 independent cohorts, HistoPLUS established superior detection and classification: | Model Architecture | Encoder Backbone | Encoder Params | Detection Quality (DQ) | Average Classification F1 | Plasmocytes F1 | Smooth Muscle F1 | Tumor Cells F1 | | -------------------- | ----------------------- | -------------- | ------------------------ | ------------------------- | -------------------------- | -------------------------- | ------------------------ | | **CellViT SAM-B** | SAM-Base | 91M | 0.716 | Baseline | 0.269 \[0.219–0.322] | 0.186 \[0.110–0.257] | 0.456 \[0.414–0.493] | | **CellViT Phikon** | Phikon (iBOT) | 86M | 0.732 | +16.5% vs SAM-B | 0.384 | 0.291 | 0.521 | | **CellViT Hibou-B** | Hibou-Base | 86M | 0.738 | +17.2% vs SAM-B | 0.395 | 0.312 | 0.534 | | **CellViT SAM-H** | SAM-Huge | 636M | 0.741 | +8.4% vs SAM-B | 0.352 | 0.264 | 0.512 | | **CellViT Virchow2** | Virchow2 (DINOv2) | 632M | 0.753 \[0.742–0.763] | +17.4% vs SAM-H | 0.461 | 0.389 | 0.579 \[0.544–0.613] | | **CellViT UNI2** | UNI2 (DINOv2) | 681M | 0.753 \[0.742–0.763] | +26.7% vs SAM-H | 0.472 | 0.398 | **0.603** \[0.570–0.634] | | **HistoPLUS (Ours)** | **H0-mini (Distilled)** | **86M** | **0.753** \[0.742–0.763] | **+36.5% vs SAM-B** | **0.483** \[0.433–0.529]\* | **0.403** \[0.282–0.515]\* | 0.583 \[0.544–0.619] | *\* Indicates statistically significant gain over baselines ($p < 0.001$).* > **Key Performance Takeaways:** > > * **Overall Metrics:** HistoPLUS delivers a **+5.2% gain in Detection Quality** and a **+23.7% gain in macro-F1 classification** over existing published baselines. > * **Understudied Cell Unlocking:** Gains are most pronounced on rare populations: plasmocytes jump from 0.269 to **0.483** (+79.5% relative), smooth muscle jumps from 0.186 to **0.403** (+116.7% relative). > * **Efficiency:** H0-mini achieves parity with 681M-parameter foundation models while consuming **$5\times$ fewer parameters** and running in less than half the wall-clock time. *** ### 2. Zero-Shot Indication Transfer (Unseen Cancers) To test cross-organ generalizability, HistoPLUS was evaluated on two cancer indications completely omitted from the training corpus: | Unseen Indication | Sample Size ($n$) | Detection Quality (DQ) | Segmentation Quality (SQ) | Lymphocytes F1 | Cancer Cells F1 | | ------------------------------------------ | ----------------- | ------------------------ | ------------------------- | ------------------------ | ------------------------ | | **Breast Invasive Carcinoma (BRCA)** | 90 ROIs | **0.836** \[0.819–0.852] | **0.801** \[0.796–0.807] | **0.799** \[0.757–0.829] | 0.475 \[0.302–0.620] | | **Ovarian Serous Cystadenocarcinoma (OV)** | 50 ROIs | **0.805** \[0.782–0.825] | **0.803** \[0.795–0.810] | **0.639** \[0.585–0.676] | **0.682** \[0.594–0.751] | Despite never encountering breast or ovarian morphology during training, HistoPLUS maintained robust boundary delineation ($SQ > 0.80$) and accurate identification of tumor-infiltrating lymphocytes ($F1 = 0.80$ in breast), proving high generalizability across novel tissue architectures. *** ## Downstream Clinical Translation: FGFR3 Mutation Prediction in Bladder Cancer Beyond cell-level benchmarking, HistoPLUS demonstrates how granular cellular characterization powers interpretable downstream clinical biomarker discovery. The authors evaluated HistoPLUS-derived cellular features for predicting **fibroblast growth factor receptor 3 (*****FGFR3*****) mutation status** in bladder urothelial carcinoma (TCGA-BLCA, $n=290$), an alteration determining patient eligibility for targeted tyrosine kinase inhibitors such as **erdafitinib**: * **Histomorphometric Feature Extraction:** Univariate logistic regression models were fitted across standardized cellular composition and spatial metrics (top 10 features ranked by cross-validated AUC). * **Immune-Cold / Desert Biological Correlation:** Consistent with known cancer biology, *FGFR3*-altered tumors exhibited a statistically significant negative association with immune cell infiltration (decreased tumor-infiltrating lymphocyte density, reduced plasmocyte infiltration, and lower stromal inflammatory index; Benjamini–Hochberg FDR $p < 0.05$). * **Clinical Triage Utility:** While next-generation sequencing and RT-PCR remain the definitive gold standard for *FGFR3* alterations, cell-level phenotyping on routine H\&E slides provides a rapid, cost-free pre-screening triage tool to prioritize biopsies for confirmatory molecular testing in resource-constrained clinical settings. *** ## Model Limitations, Class Confusion & Diagnostic Safeguards Granular error analysis from the multi-cohort confusion matrix (Supplementary Fig. 4) highlights critical areas for caution and human-in-the-loop oversight: 1. **Detection Dropouts in Minute or Subtle Structures:** * Red blood cells, apoptotic bodies, neutrophils, and macrophages exhibit the highest dropout rates. Small, fragmented pyknotic chromatin spheres (apoptotic bodies) can be overlooked in basophilic cellular debris. 2. **Stromal Lineage Phenotypic Confusion:** * Spindle-shaped mesenchymal elements (fibroblasts, smooth muscle cells, and vascular endothelial cells) share elongated nuclear morphology and fibrillar eosinophilic stroma, representing the primary axis of classification confusion. 3. **Cytologic Mitotic / Apoptotic Overlap:** * Early prophase or atypical condensed mitoses can occasionally be misattributed to apoptotic bodies due to condensed, dark chromatin without an intact nuclear envelope. 4. **Clinical Deployment Guardrails:** * The authors emphasize that HistoPLUS is designed as an **assistive diagnostic and research engine** rather than an autonomous sign-out tool. Clinical application requires pathologist audit and task-specific fine-tuning or threshold calibration. *** ## Code & Practical Workflows ### 1. Installation ```bash # System dependencies (Ubuntu/Debian) sudo apt-get install openslide-tools # Install directly from GitHub git clone https://github.com/owkin/histoplus.git cd histoplus pip install -e . ``` ### 2. High-Throughput WSI Inference (CLI) ```bash # Run cell detection, segmentation, and typing across whole slides histoplus \ --slides /path/to/slide.svs \ --export_dir ./predictions/ \ --batch_size 16 ``` ### 3. Native Python Inference API ```python import openslide from histoplus.extract import extract from histoplus.helpers.hub import histoplus_cellvit_segmentor_40x from histoplus.helpers.tiling import get_wsi_coordinates # 1. Open Whole Slide Image slide_path = "sample_colon_wsi.svs" slide = openslide.OpenSlide(slide_path) # 2. Load Pretrained 40x Segmentor (H0-mini ViT-Base Backbone, MPP 0.25) segmentor = histoplus_cellvit_segmentor_40x(inference_image_size=784) # 3. Extract Grid Coordinates across Tissue Regions coords, deepzoom_level = get_wsi_coordinates(slide, mpp=0.25, tile_size=784) # 4. Extract Instance Masks & 13-Class Annotations segmentation_results = extract( slide=slide, coords=coords, deepzoom_level=deepzoom_level, segmentor=segmentor, tile_size=784, batch_size=16, inference_tile_overlap=64, # Prevents border truncation n_workers=4, ) # 5. Export to GeoJSON for QuPath segmentation_results.to_geojson("./qupath_annotations.geojson") ``` *** ## Vault Context & Comparative Positioning Within the evolving computational pathology ecosystem, HistoPLUS occupies a complementary role alongside existing single-cell and foundation model architectures: 1. **HistoPLUS vs. \[\[CytoFormer: A Molecularly Supervised Cell Foundation Model for Histopathology Cell Classification]]:** * *Supervision Paradigm:* CytoFormer uses automated spatial transcriptomics (10x Xenium) to eliminate human annotations entirely across 15.4M cells, while HistoPLUS uses active-learning-guided multi-pathologist consensus (HistoVAL) combined with \[\[NuClick]] boundary propagation across 108k nuclei. * *Architecture & Routing:* CytoFormer routes single-cell crops through 16 organ-specific linear heads attached to a ViT-giant encoder; HistoPLUS uses an all-in-one CellViT with a compact 86M distilled foundation model (H0-mini) that performs simultaneous boundary segmentation and instance classification without requiring organ routing signals. 2. **Complementarity with Synthetic Benchmarks (\[\[HistoGen: Histopathology Cell Nuclei Image Generation Tool]]):** * HistoGen, developed by FDA CDRH/DIDSR, generates synthetic nuclei conditioned on HoVer-Net horizontal/vertical distance maps to stress-test digital pathology algorithms. HistoPLUS provides the inverse clinical engine: parsing real-world WSIs, generating precise HV maps and instance boundaries, and classifying 13 native cell lineages. 3. **Foundation Model Trade-Offs (\[\[Navigating foundation model selection in digital pathology through performance evaluation and tradeoff analysis]]):** * Validates the core thesis established by Trager et al.: larger parameter counts (UNI2, Virchow2) do not equate to superior cellular performance. HistoPLUS proves that an 86M distilled model trained on diverse, active-learning-enriched morphology matches or surpasses 600M+ models on cell-level tasks while slashing runtime by more than half. 4. **Integration with QuPath & Spatial Toolchains:** * With native GeoJSON export, HistoPLUS outputs can be directly imported into \[\[From Samples to Knowledge 2025: QuPath Training Course]] and \[\[Multiplex Immunofluorescence Image Analysis with QuPath — Part 1: Understanding Digital Images]] to construct cell graphs, analyze spatial proximity distributions, and quantify localized tumor-immune niches. *** ## Related Notes * **Cell Segmentation & Typing:** \[\[NuClick]], \[\[CytoFormer: A Molecularly Supervised Cell Foundation Model for Histopathology Cell Classification]], \[\[HoVer-NeXt]], \[\[HistoGen: Histopathology Cell Nuclei Image Generation Tool]], \[\[Micro-Manager]] * **Foundation Models & Trade-Offs:** \[\[Navigating foundation model selection in digital pathology through performance evaluation and tradeoff analysis]], \[\[Towards robust foundation models for digital pathology]], \[\[A distributional robustness margin for pathology foundation models]] * **Slide-Level Aggregators & MIL:** \[\[A Hybrid MIL Approach Leveraging Convolution and State-Space Model for Whole-Slide Image Cancer Subtyping]], \[\[Weakly supervised MIL histopathological tumor segmentation]], \[\[TRICARE: Deep-learning triage of 3D pathology datasets]] * **Spatial Histology & Workflows:** \[\[Tumor budding T-cell graphs for pT1 colorectal cancer]], \[\[Multiplex Immunofluorescence Image Analysis with QuPath — Part 1: Understanding Digital Images]], \[\[From Samples to Knowledge 2025: QuPath Training Course]] * **Standards & Catalogs:** \[\[Regulatory Science Tools Catalog: Digital Pathology (FDA CDRH)]], \[\[Considerations for digital pathology displays]], \[\[What AI Can and Cannot Do in Pathology]]