Loading…
Histopathology tissue detection model that segments whole-slide thumbnails in a single SAM2 forward pass, replacing patch-wise slide preprocessing.
No providers recorded yet. Browse all providers
A diagnostic whole-slide image is mostly glass. Before any pathology model sees a single cell, something has to decide which of the gigapixel's pixels are tissue, and that decision sets the search space for everything downstream: miss tissue and the information is gone, keep background and every later stage pays in storage, I/O and encoder time. Colour thresholding is fast but brittle — it breaks on stain variation, faint biopsies, pen markings and scanner streaks. Deep tissue detectors are more robust but run at the patch level, spending hundreds to thousands of forward passes on a slide just to find out where the slide is.
AtlasPatch moves detection to the coarsest place it can live. Its tissue detector reads the slide's 1.25× thumbnail once, returns a binary tissue-versus-background mask, and the pipeline extrapolates that mask's contours back into full-resolution slide coordinates to generate patch locations at whatever magnification the downstream task wants. No patch is ever read to decide whether a patch should be read. The detector is a parameter-efficient adaptation of SAM2, Meta's promptable segmentation foundation model: the image encoder and all attention and MLP weights stay frozen, and only the affine parameters of the layer-normalization layers are updated — 0.076% of the model's weights.
Released in February 2026 by the Atlas Analytics Lab at Concordia University, with collaborators at the CHUM, Mila, McGill and Khalifa University, it sits beside SAM-family adaptations to histology such as PathoSAM but works at a different level: where those segment nuclei inside a tile, AtlasPatch decides which tiles exist at all.
The detector is SAM2.1 Hiera-Tiny, a 38.9M-parameter model whose hierarchical vision transformer backbone is roughly 6× faster at inference than the original SAM's ViT encoder. It runs image-only, with no memory bank or temporal propagation, prompted with a full-frame bounding box so the mask decoder gets global spatial context without manual prompting. Training used thumbnails resized to 1024×1024, a weighted loss of 0.65·Dice plus binary cross-entropy, AdamW at 5×10⁻⁴, batch size 2, and 50 epochs with early stopping on validation F1. The corpus was drawn from 35,827 WSIs across 11 datasets — in-house pancreas and digestive-system cohorts from the CHUM, plus CAMELYON17, PANDA and eight TCGA projects — and labelled by three trained annotators under quality-control review, with a board-certified pathologist vetting the result.
On a held-out test set of 3,000 slides the detector reaches a tissue-detection precision of 0.986, ahead of Trident-Hest (0.983), HistoQC (0.985) and dplabtools (0.977), varying by under 4% across stratified subsets of brightness, fragmentation and tissue percentage. Detection runs 2.6× faster than Trident-GrandQC and 20× faster than Trident-Hest, and the full pipeline extracts patch coordinates for 100 slides in 195.5 seconds. Across six slide-level MIL tasks — colorectal invasiveness and dysplasia, TCGA-BRCA and TCGA-KIRC-versus-KIRP subtyping, lung adenocarcinoma versus squamous carcinoma, and PANDA Gleason grading, all embedded with UNI — the resulting patches match or beat the alternatives while yielding roughly 3,047 patches per slide against CLAM's 8,976.
The immediate beneficiaries are groups assembling slide corpora at foundation-model scale, where preprocessing dominates wall-clock time and a brittle detector silently corrupts the dataset: CLAM's heuristic masking finds no tissue at all in roughly 54% of PANDA prostate biopsies. AtlasPatch ships as a Python package and command-line tool over OpenSlide-backed slide formats, with a registry of 65+ patch encoders, so a lab can stop at tissue masks, at patch coordinates, or at ready-to-train embeddings — or reuse its mask overlays for clinical quality control.
AtlasPatch addresses a step usually treated as plumbing, and argues that the plumbing has become the constraint: corpora of 100,000-plus slides make repeated patch-level inference a real bottleneck. That a frozen general segmentation backbone can be moved into histopathology by training 0.076% of its weights is useful beyond this task. Three limitations are worth stating: the paper is a preprint under review with no journal version; the code and the released checkpoint both carry a CC BY-NC-SA 4.0 licence excluding commercial use; and the in-house CHUM cohorts supplying most of the training corpus are not publicly available, so the annotated masks were not released.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.