bio.rodeo
ModelsOrganizationsProvidersLeaderboardAboutSign in
bio.rodeo

The authoritative source for evaluating biological foundation models. No hype, just honest analysis.

Categories
  • DNA & Gene
  • RNA
  • Protein
  • Small molecule
  • Single-cell
  • Spatial omics
  • Pathology
  • Imaging
  • Metabolomics
  • Biosignals
  • Language model
bio.rodeoModelsOrganizationsProvidersLeaderboardAboutFAQSubmit a modelContact
© 2026 Pulsatance. All rights reserved. ~
Built by Pulsatance
models / pathology / iqmai
PathologyImaging
PathAIReleased March 2025

iQMAI

Digital pathology model inferring collagen structure from H&E whole-slide images, trained against polarization imaging of Picrosirius Red slides.

The short version

  • —Reads collagen off routine H&E slides, with no special stain or extra glass required
  • —Extracts per-fiber width, length, tortuosity, and angle to the epithelial-stromal interface
  • —Ground truth comes from polarization imaging of the same slide, registered pixel to pixel
  • —One checkpoint spans lung, liver, and pancreatic cohorts without per-cohort retraining
  • —Collagen overlays pair with tissue and cell segmentation for spatially resolved features
7Openness3Citations
0GitHub stars

Where to run it

No providers recorded yet. Browse all providers

Fibrillar collagen is the most abundant component of the tumor extracellular matrix, and its organization — fiber width, length, tortuosity, and alignment relative to the tumor edge — carries prognostic weight in breast, lung, head and neck, and pancreatic cancer. Seeing it has always required something other than the routine hematoxylin and eosin (H&E) slide: a special stain such as Picrosirius Red (PSR), or second-harmonic generation microscopy on an unstained section. Special stains vary between laboratories and the microscopy needs bespoke optics, so collagen is rarely quantified at whole-slide scale.

iQMAI — inferred quantitative multimodal anisotropy imaging — is PathAI's approach to recovering that signal from the slide pathologists already have. Because individual collagen bundles cannot be reliably annotated by eye, the training labels were generated physically rather than manually. H&E slides were scanned, de-stained, re-stained with PSR, and imaged on a custom microscope built around two cross-polarized linear polarizers on a motorized rotational stage. The resulting polarization image, which reports tissue birefringence, was registered pixel-to-pixel back onto the H&E image through a multispectral channel, giving an exhaustive "super-annotation" map from which the model learns to predict polarization signal directly from H&E.

It sits alongside virtual-staining models that translate H&E into immunohistochemistry, such as UNIStainNet, but the target is a physical optical measurement rather than an antibody readout, and the product is measured fiber geometry rather than a synthetic image. Posted in March 2025, it was deployed as a single checkpoint across four TCGA cohorts, with no per-cohort retraining.

#Key Features

  • Collagen from a routine slide: Inference runs on the standard H&E whole-slide image, so no additional glass, special stain, or birefringence-sensitive hardware is required.
  • Physically generated ground truth: Labels come from PSR-enhanced polarization imaging of the same tissue section, sidestepping the impossibility of hand-annotating individual fibers.
  • Per-fiber morphology: A modified FIRE fiber-extraction algorithm converts the predicted collagen map into individual fibers with width, length, tortuosity, and angle relative to the epithelial-stromal interface, aggregated to patch and slide level.
  • Scanner robustness: Collagen intensity agreed closely across four commercial scanners (AT2, DP200, GT450, C13220), with intraclass correlations of 0.92 to 0.98 against AT2.

#Technical Details

The network is a U-Net used for dense regression, with six downsampling and six upsampling levels stacking ConvNormAct blocks — 2D convolution, batch normalization, SiLU activation. Only valid convolutions are used, so a 1020 x 1020 pixel RGB H&E input yields a 260 x 260 pixel intensity map free of patch-boundary artifacts. Training minimized a weighted sum of LPIPS perceptual loss and a custom differentiable DICE loss (0.01 and 0.99 respectively), the latter added to correct undercall where polarization signal is strong. The development set comprised 337 slides across 15 cancer indications, split into 244 training, 47 validation, and 46 held-out test slides, yielding 4,926,844 training and 666,095 validation frame pairs. Optimization used AdamW on 32 RTX 3080 GPUs for 500,000 iterations at a learning rate of 5e-4 and a minibatch size of 2560, with ContriMix stain-color augmentation.

Against ground-truth polarization images, patch-level agreement reached a mean structural similarity index of 0.84, a mean patch-wise RMSE of 0.04, a linear fit of R2 = 0.93, and an intraclass correlation of 0.96. Fiber-level agreement was strongest for median width and relative angle, weaker for length and tortuosity. In a pathologist review of 100 test frames, all showed 0-10% undercall and 71% showed under 30% overcall.

#Applications

The trained model was deployed on TCGA H&E slides from lung adenocarcinoma (n=492), lung squamous cell carcinoma (n=416), hepatocellular carcinoma (n=369), and pancreatic ductal adenocarcinoma (n=204), combined with tissue and cell segmentation to yield spatially resolved features. Collagen fiber density tracked fibroblast density more tightly in pancreatic (Spearman rho = 0.59) and hepatocellular (0.64) tumors than in non-small cell lung cancer (0.32), and fiber width, tortuosity, and relative angle all increased with distance from the epithelial-stromal interface. In pancreatic cancer, fiber density and median width were negatively associated with the immunosuppressive LRRC15-CAF gene expression signature, and wider fibers were associated with longer overall survival (n=132). The beneficiaries are translational and biomarker teams wanting matrix organization as a covariate on H&E-only cohorts.

#Impact

iQMAI removes the practical barrier that has kept collagen organization out of large retrospective studies, and because it runs on the same image as tissue and cell models, it puts matrix structure into the same spatial frame as the rest of the microenvironment. The caveats are concrete. Other birefringent materials — keratin, fibrin, surgical ink — and eosinophilic fibrous tissue such as smooth muscle produce false-positive signal, mitigated by masking predictions to cancer stroma. Fiber length is systematically overcalled where fibers leave the plane of a two-dimensional section. Roughly 90% of training slides came from a single laboratory, a color-diversity limitation addressed through stain augmentation rather than broader sourcing. The work remains a preprint, and neither model code nor weights have been released: the repository contains only the notebooks generating two of the paper's figures, under a custom license.

At a glance

Released
March 2025
Category
Pathology
License
NOASSERTION
Organization
PathAI

Related models

  • MIPHEI-ViT

    Sanofi / Mines Paris – PSL / Institut Curie

  • HistDiST

    University of Tübingen

  • Path2Omics

    National Cancer Institute / Australian National University / Sungkyunkwan University / University of Pennsylvania / Cedars-Sinai Medical Center

  • SMILE

    Johns Hopkins University

  • PathQC

    Sanford Burnham Prebys

Links

GitHub RepositoryResearch Paper

Tags

histologytumor_microenvironmentu_netvirtual_staining

Something wrong?

Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.