PLUTO-4

PLUTO-4 is a family of self-supervised vision transformer foundation models for digital pathology developed by PathAI and described in an arXiv preprint posted in November 2025 (lead author Harshith Padigela with thirteen co-authors). The models are general-purpose feature extractors for histopathology, learning rich morphological representations from whole-slide images without manual annotation, and serving as backbones for downstream tile classification, cell and tissue segmentation, and slide-level diagnosis.

PLUTO-4 is the frontier-scale successor to the original PLUTO (Pathology-Universal Transformer, arXiv:2405.07905), a compact ~22M-parameter ViT-S model. Whereas the original PLUTO emphasized efficiency and multi-task adaptability at small scale, PLUTO-4 retains that compact-deployment lineage in one variant while introducing a second, dramatically larger model aimed at maximizing representational capacity. The release thus spans two distinct points on the efficiency-versus-capacity tradeoff rather than replacing the original outright.

The work targets a recurring problem in computational pathology: no single backbone has reliably generalized across the full diversity of tissue types, disease indications, and staining protocols seen in real clinical practice. PLUTO-4 addresses this by scaling both model size and the breadth of pretraining data, reporting new state-of-the-art results across multiple pathology benchmarks, including an 11% improvement in dermatopathology diagnosis from the flagship model.

#Key Features

• Two complementary variants: PLUTO-4S is a compact ~22M-parameter model tuned for cost-efficient, multi-scale deployment, while PLUTO-4G is a ~1.1B-parameter frontier-scale model built to maximize representation capacity for the most demanding tasks.
• Multi-scale FlexiViT design (PLUTO-4S): The compact variant uses a FlexiViT setup with 2D rotary position embeddings (2D-RoPE), allowing a single model to operate flexibly across multiple patch sizes and magnifications.
• Highly diverse pretraining corpus: Training spans 551,164 whole-slide images from 137,144 patients across 50+ institutions, covering 60+ disease types and 100+ stains, including H&E, immunohistochemistry, and special stains.
• Self-supervised pretraining: Both models are pretrained with a DINOv2-derived self-supervised objective, learning generalizable tissue representations without labeled data.
• Broad task coverage: A single backbone supports tile classification, cell and tissue segmentation, and slide-level diagnosis, with reported first-in-class results in cell segmentation and spatial transcriptomics prediction.

#Technical Details

PLUTO-4 extends the original PLUTO (Pathology-Universal Transformer) lineage with vision transformer backbones pretrained using a self-supervised objective derived from DINOv2. The two variants make opposite architectural choices to suit their goals. PLUTO-4S, at roughly 22 million parameters, adopts a FlexiViT configuration with 2D-RoPE position embeddings so it can ingest variable patch sizes and operate efficiently across magnifications for multi-scale deployment. PLUTO-4G, at roughly 1.1 billion parameters, instead trains with a single fixed patch size to concentrate capacity on maximizing representational quality. The shared pretraining dataset comprises 551,164 whole-slide images drawn from 137,144 patients across more than 50 institutions, spanning over 60 disease types and more than 100 staining protocols — a corpus markedly more diverse in stains and indications than earlier PLUTO releases.

On evaluation, PLUTO-4G establishes new performance frontiers across multiple pathology benchmarks, with reported first-in-class results in cell segmentation and spatial transcriptomics prediction and an 11% improvement on a complex 17-class internal dermatopathology diagnosis benchmark relative to the prior PLUTO generation.

#Applications

PLUTO-4 functions as a feature backbone for building downstream computational pathology systems rather than as an end-to-end diagnostic product. Researchers and pathology AI teams extract patch-level embeddings and aggregate them — for example with multiple-instance learning — to drive cancer detection, tissue and tumor subtyping, histological grading, and biomarker inference directly from morphology. The compact PLUTO-4S variant suits cost-sensitive or high-throughput deployments and multi-magnification pipelines, while the larger PLUTO-4G is intended for use cases that demand maximum accuracy, such as cell segmentation, spatial transcriptomics prediction, and fine-grained dermatopathology classification. The broad multi-stain pretraining makes the models applicable beyond H&E to immunohistochemistry and special-stain workflows where H&E-only backbones tend to underperform.

#Impact

PLUTO-4 reflects the field's continued shift from narrow, task-specific pathology models toward general-purpose foundation backbones, and demonstrates that scaling both model size and stain/indication diversity yields measurable gains on demanding clinical tasks. By releasing a paired compact and frontier-scale family, PathAI offers practitioners an explicit capacity-versus-cost choice from a shared pretraining regime, with the flagship feeding internal products such as PathAssist Derm. Important caveats apply: as of the November 2025 preprint, model weights and code do not appear to be publicly released, and benchmark results — including the dermatopathology improvement — derive substantially from internal evaluations that warrant independent confirmation. Like other pathology foundation models, PLUTO-4 is a research tool and has not received regulatory clearance for clinical diagnostic use.