CanvOI

Digital pathology foundation models inherited their input geometry from natural-image vision transformers: 224 x 224 pixel tiles cut into 14- or 16-pixel patches. A whole-slide image is not a photograph, though. Downstream, an attention-based multiple instance learning (MIL) aggregator folds thousands of tile embeddings into one slide-level call, and every additional tile is one more vector it has to weigh. CanvOI's contribution is to change that geometry rather than the parameter budget: it reads 380 x 380 pixel tiles cut into 10 x 10 pixel patches. The larger tile summarizes more tissue per embedding, so a slide yields fewer vectors for the aggregator; the smaller patch lets the transformer resolve finer structure inside each one.

CanvOI — "Cancer Vision Oncology Intelligence" — is a 1.1 billion parameter ViT-g/10 pretrained with DINOv2 self-supervision, posted to arXiv in September 2024 by Imagene AI with co-authors at Oracle and the Ellison Institute of Technology. It targets oncologic histopathology specifically rather than general histology, and runs on Oracle Cloud Infrastructure.

The comparison the paper sets up is against the frozen-encoder pathology models released earlier that year — H-optimus-0, Prov-GigaPath, Virchow and Hibou-L — all of which operate on 224-pixel tiles. The first two are also ViT-g models at the same 1.1B scale, which is what isolates tile and patch size as the variable under test.

#Key Features

• Input geometry as a scaling axis: Tile size and patch size are treated as tunable alongside data volume and parameter count, rather than carried over from ImageNet-era conventions.
• Built for the MIL bottleneck: Fewer, richer embeddings per slide are easier for an attention-based aggregator to combine than many thin ones.
• Frozen encoder, task-specific head: Only the AB-MIL head is trained per task, so a new question costs a small labeled cohort rather than a new model.
• Low-label performance: The margin over comparator encoders widens as labeled slides are removed, which matters for rare conditions and small clinical-trial cohorts.
• Pan-cancer, multi-site pretraining: More than 40 organs and tissue types drawn from over 100 international source sites.

#Technical Details

Pretraining used 70,217,688 tissue tiles from 632,608 tissue samples, almost exclusively hematoxylin and eosin stained and all normalized to 20x magnification (0.5 µm/pixel). More than 90% came from Imagene's internal holdings; the remainder from TCGA, CPTAC, GTEx, the Cancer Moonshot Biobank and CAMELYON17. Tissue images were extracted at 670 x 670 pixels, from which global views (crop range 0.2–0.57) were resized to 380 x 380 and local views (0.03–0.2) to 140 x 140 for the DINOv2 teacher-student scheme. To control cost, the model was first trained at 224 x 224 and then fine-tuned at 380 x 380. One tile costs 3,800 GFLOPs against 625 for H-optimus-0, 475 for Prov-GigaPath, 367 for Virchow and 182 for Hibou-L.

Evaluation was slide-level only, through AB-MIL heads trained identically across encoders (Adam, learning rate 1e-4, weight decay 1e-5, 15 epochs). On BRACS breast tissue (547 WSIs from 189 patients) CanvOI reached 0.898 AUC on three-class lesion type and 0.812 on the seven-class subtype task; on HunCRC colorectal tissue, 0.960 on four-class lesion type; and on an internal 1,079-slide NSCLC cohort held out by source site, 0.962 on histological subtype and 0.903 on biopsy site. Averaged over the five tasks that is 0.905, against 0.890, 0.887, 0.881 and 0.831 for the four comparators. Cutting the labeled lung cohort to 10% — roughly 100 slides — left CanvOI at 0.83 average AUC, its widest margin of any label budget tested.

#Applications

CanvOI is a backbone rather than a diagnostic: the intended use is to embed whole-slide images once, then train small heads for the questions an oncology group actually asks — biomarker discovery, treatment-response prediction, prognosis, and subtype or biopsy-site classification. Its behavior under label scarcity suits translational and clinical-trial work, where a cohort may run to tens or low hundreds of annotated cases. Access is through Imagene's OI Suite, a hosted platform on Oracle Cloud Infrastructure; there is no public checkpoint, so it fits groups willing to work in a vendor environment rather than those needing weights on their own hardware.

#Impact

The paper's value is the question it reframes. Pathology foundation models had been scaled along two axes — corpus size and parameter count — and CanvOI argues for a third, showing that redistributing the same 1.1B parameters across a different tile-and-patch decomposition buys 1.5–7.4% averaged AUC and a larger gain still in the low-label regime. The evidence has real limits: it remains a preprint, two of the five tasks run on an unreleased internal lung cohort, and each comparator was run at its own native 224-pixel tile size, so encoder and input geometry move together rather than independently. Neither the weights, the training code, nor the pretraining corpus is public, and there is no model card, so the result cannot be reproduced outside Imagene's platform.