Prov-GigaPath: ViT Digital Histopathology Model

• Prov-GigaPath is a pathology-specific foundation model utilizing a ViT-based patch encoder and supplementary slide-level encoder for whole-slide image analysis.
• It is pre-trained on approximately 1.3 billion image patches from over 171,000 WSIs, capturing diverse tissue types and supporting tasks like LUAD pattern prediction and zero-shot retrieval.
• The model's versatility is enhanced through fine-tuning, adapter-based multimodal fusion, and attention-driven feature aggregation, while its limitations include sensitivity to domain shifts and computational demands.

Prov-GigaPath is a pathology-specific foundation model family for digital histopathology and whole-slide image analysis. Across the published literature, it is described as a ViT-G or ViT-G/14 model trained with DINOv2-style self-supervision on approximately 1.3 billion pathology image patches from 171,189 whole-slide images (WSIs) from the Providence Health System, with several papers distinguishing a high-capacity patch encoder from an additional slide-level encoder used for WSI modeling (Perez-Herrera et al., 23 Apr 2026, Mishra et al., 18 Sep 2025). The literature uses the names “Prov-GigaPath,” “Prov-Gigapath,” and “GigaPath” for this Providence model family, and evaluates it as a frozen representation learner, a fine-tuned patch encoder, a visual backbone for vision–language adaptation, and a component in multi-model fusion pipelines (Alfasly et al., 2024, Flannery et al., 11 Dec 2025).

1. Model definition and architectural scope

Prov-GigaPath is consistently presented as a very large ViT-based pathology foundation model. In the lung adenocarcinoma study, its patch encoder is specified as a ViT-G/14 with 1.1B parameters, and the full architecture includes an additional slide-level encoder, although only the patch encoder was used in that work. In the cell segmentation and classification study, it is instantiated as a Vision Transformer with 40 transformer blocks, input resolution 224×224, patch size 14×14, 1.141 billion parameters, and FLOPs ranging from 78 GFLOPs to 303 GFLOPs depending on which blocks are tapped for features (Perez-Herrera et al., 23 Apr 2026, Vadori et al., 4 Feb 2025). This suggests a common ViT-G family backbone reported under study-specific implementation conventions.

A recurrent distinction in downstream usage is between the patch encoder and the WSI-level model. In MGPATH, Prov-GigaPath is treated strictly as a frozen patch encoder $E_I(\cdot)$ that outputs 1536-dimensional tile embeddings, while slide-level aggregation is replaced by multi-granular prompting and optimal-transport-based visual–text matching. In zero-shot retrieval, by contrast, one evaluation uses the native slide-level embedding of “GigaPath WSI,” whereas another uses patch embeddings from the same family inside Yottixel’s patch-based search pipeline (Nguyen et al., 11 Feb 2025, Alfasly et al., 2024).

2. Pretraining corpus and representation geometry

The model’s scale is one of its defining characteristics. Multiple studies describe pretraining on approximately 1.3 billion or nearly 1.4 billion patches derived from 171,189 WSIs, with H&E and IHC represented across 31 tissue types and more than 30,000 patients. One evaluation specifies 256×256 patches at 20× magnification, and the lung adenocarcinoma study emphasizes that the pretraining corpus is rich in lung content, with Lung proportion 45.29%, which is especially relevant for LUAD (Vadori et al., 4 Feb 2025, Perez-Herrera et al., 23 Apr 2026). In practical downstream pipelines, this pretraining has been leveraged both in frozen form and after limited task-specific adaptation.

Representational analyses indicate that scale and pathology specificity do not translate into a simple notion of uniqueness. Using representational similarity analysis on 50,000 TCGA H&E patches from BRCA, LUAD, LUSC, and COAD, Prov-GigaPath had the highest average similarity across models, with mean Spearman correlation 0.523, and its strongest pairwise similarity was with UNI2 at $\rho_S = 0.555$ with range $[0.531, 0.575]$ across batches. At the same time, it showed the highest slide specificity among the compared computational pathology models, with $\delta_{\text{slide}} = 0.762$ pre-normalization and $0.712$ after Macenko normalization, while disease specificity remained low at $\delta_{\text{disease}} = 0.138$ pre-normalization and $0.124$ post-normalization. Spectral analysis placed Prov-GigaPath among the models with the most shallow cumulative singular-value rise, indicating a more distributed, higher-dimensional representational structure (Mishra et al., 18 Sep 2025). This suggests that the model occupies a broadly shared pathology embedding space while retaining strong slide-specific signatures.

3. Weakly supervised predominant growth pattern prediction in LUAD

A detailed pathology-task evaluation of Prov-GigaPath appears in attention-based multiple instance learning for predominant growth pattern prediction in lung adenocarcinoma WSIs. In that study, Prov-GigaPath’s role was strictly as a patch encoder; its original whole-slide encoder was not used. Patch-level domain adaptation was performed on ANORAK, comprising 731 ROIs from 49 H&E WSIs annotated at approximately $0.45\ \mu$m/pixel (20×), from which 4,085 non-overlapping 448×448 patches were extracted, split at patient level into 3,226 training and 859 validation patches. Slide-level training and evaluation used DHMC, a set of 143 H&E WSIs labeled by consensus of three pathologists for predominant pattern, with five-fold cross-validation because of limited sample size (Perez-Herrera et al., 23 Apr 2026).

Within the ABMIL pipeline, each Prov-GigaPath patch embedding was projected through a linear layer into a 512-dimensional representation before class-specific gated attention pooling. The attention weights were written as

$$a_k^{(c)} = \frac{\exp\left\{ \mathbf{w}^{(c)} \left[ \tanh(\mathbf{V}\mathbf{h}_k) \odot \sigma(\mathbf{U}\mathbf{h}_k) \right] \right\}}{\sum_{j} \exp\left\{ \mathbf{w}^{(c)} \left[ \tanh(\mathbf{V}\mathbf{h}_j) \odot \sigma(\mathbf{U}\mathbf{h}_j) \right] \right\}},$$

with class-specific bag representations

$$\mathbf{z}^{(c)} = \sum_{k} a_k^{(c)} \mathbf{h}_k,\quad y^{(c)} = \mathbf{w}_{\text{cls}^{(c)}} \mathbf{z}^{(c)}.$$

Fine-tuning on ANORAK improved already strong patch-level agreement, and Prov-GigaPath reached the highest validation agreement at $\rho_S = 0.555$0. At the slide level, majority voting aggregation produced weaker results than ABMIL. Under frozen ABMIL, Prov-GigaPath achieved weighted F1 $\rho_S = 0.555$1 and $\rho_S = 0.555$2; under fine-tuned ABMIL, it was best overall at weighted F1 $\rho_S = 0.555$3 and $\rho_S = 0.555$4, which is the abstract’s stated highest agreement. These results significantly outperformed all other models except UNIv1. Prov-GigaPath also exhibited the most balanced per-class behavior and achieved non-zero F1-scores for all subtypes, although papillary remained difficult with $\rho_S = 0.555$5 slides and F1 $\rho_S = 0.555$6 across models. The paper further reports that ABMIL attention maps localized prognostically relevant LUAD patterns across WSIs, supporting slide-level interpretability. A plausible implication is that Prov-GigaPath’s lung-heavy pretraining corpus interacts favorably with weakly supervised slide-level attention in LUAD.

4. Frozen-encoder transfer in cell analysis and zero-shot retrieval

Prov-GigaPath has also been evaluated as a frozen encoder in cell instance segmentation and classification. In that setting, it serves as one of three histopathology-specific ViT backbones inside a common UNETR-inspired encoder–decoder pipeline derived from CISCA, with features extracted from four encoder depths via skip connections. For Prov-GigaPath, three block-selection strategies were tested on PanNuke: shallow $\rho_S = 0.555$7, deep $\rho_S = 0.555$8, and mixed $\rho_S = 0.555$9. The mixed configuration was clearly best, reaching $[0.531, 0.575]$0, $[0.531, 0.575]$1, $[0.531, 0.575]$2, $[0.531, 0.575]$3, $[0.531, 0.575]$4, and $[0.531, 0.575]$5, while the shallow variant retained the highest $[0.531, 0.575]$6. Even so, Prov-GigaPath trailed UNI2 and Virchow2 among histopathology ViTs, and general-purpose non-ViTs such as Swin2-B-22K and ConvNeXt-B-22K outperformed all histopathology ViTs on this task. The study also reports higher training and validation losses for Prov-GigaPath and signs of underfitting, arguing that locality and hierarchical inductive biases matter more than domain-specific pretraining for small-cell instance tasks unless ViT features are carefully tapped (Vadori et al., 4 Feb 2025). A common assumption that pathology-specific pretraining necessarily dominates general-purpose vision pretraining is therefore not supported for frozen CISCA.

In zero-shot whole-slide image retrieval, Prov-GigaPath appears in both patch-based and native WSI-level forms. Under Yottixel-GigaPath, patch embeddings from the patch backbone are indexed through mosaic sampling, barcode compression, and median-of-minimums Hamming distance; under GigaPath WSI, the model’s own slide-level embedding is used directly. Overall macro-average F1 was similar for the two modes: top-1 $[0.531, 0.575]$7 versus $[0.531, 0.575]$8, majority top-3 $[0.531, 0.575]$9 versus $\delta_{\text{slide}} = 0.762$0, and majority top-5 $\delta_{\text{slide}} = 0.762$1 versus $\delta_{\text{slide}} = 0.762$2 for 17 organs. These numbers were competitive with Yottixel-UNI at $\delta_{\text{slide}} = 0.762$3 and Yottixel-Virchow at $\delta_{\text{slide}} = 0.762$4, and substantially above Yottixel-DenseNet at $\delta_{\text{slide}} = 0.762$5, but the paper emphasizes that absolute performance remained modest. Organ-wise performance was heterogeneous, with stronger results in esophagus, kidneys, bladder, colon, and rectum, and much lower results in lungs and skin (Alfasly et al., 2024). This suggests that Prov-GigaPath embeddings are useful for archive-scale similarity search, but not sufficient for reliable subtype-level retrieval without further task alignment or domain adaptation.

5. Multimodal adaptation and information-driven fusion

Prov-GigaPath has been adapted beyond pure visual encoding. MGPATH extends the frozen Prov-GigaPath visual encoder into a vision–LLM by pairing it with the PLIP text encoder through trainable MLP adapters $\delta_{\text{slide}} = 0.762$6 and $\delta_{\text{slide}} = 0.762$7 and contrastive learning on approximately 923,000 pathology image–text pairs from ARCH, PatchGastricADC22, and Quilt-1M. In that configuration, the visual backbone remains frozen and outputs 1536-dimensional tile embeddings, the PLIP text encoder remains frozen and outputs 512-dimensional text embeddings, and only the adapters are trained, yielding approximately 5.35M trainable parameters for the GigaPath–PLIP configuration. Slide-level modeling then uses multi-granular prompt learning, in which learnable visual prompt tokens attend both to individual patch embeddings and to spatially grouped tokens, and classification is performed with optimal transport or unbalanced optimal transport rather than cosine similarity. The resulting PLIP-G variant improved consistently over PLIP-only on few-shot WSI classification: on TCGA-NSCLC it reached AUC $\delta_{\text{slide}} = 0.762$8, F1 $\delta_{\text{slide}} = 0.762$9, and ACC $0.712$0; on TCGA-RCC, AUC $0.712$1, F1 $0.712$2, and ACC $0.712$3; and on TCGA-BRCA, AUC $0.712$4, F1 $0.712$5, and ACC $0.712$6 (Nguyen et al., 11 Feb 2025). The reported ablations further indicate that trained adapters outperform random adapters and that directly tuning Prov-GigaPath layers is not necessary and can reduce performance versus adapter-only alignment.

A separate line of work studies Prov-GigaPath as one component of tile-level foundation-model fusion. In a CLAM-based framework for kidney grading, prostate grading, and rectal staging, Prov-GigaPath alone achieved hold-out test performance of AUC $0.712$7, Sens $0.712$8, Spec $0.712$9, and F1 $\delta_{\text{disease}} = 0.138$0 on kidney; AUC $\delta_{\text{disease}} = 0.138$1, Sens $\delta_{\text{disease}} = 0.138$2, Spec $\delta_{\text{disease}} = 0.138$3, and F1 $\delta_{\text{disease}} = 0.138$4 on prostate; and AUC $\delta_{\text{disease}} = 0.138$5, Sens $\delta_{\text{disease}} = 0.138$6, Spec $\delta_{\text{disease}} = 0.138$7, and F1 $\delta_{\text{disease}} = 0.138$8 on rectal. Although it was not the top-performing single FM on prostate or rectal, including it within correlation-guided intelligent fusion improved or matched the best single-FM results: kidney reached F1 $\delta_{\text{disease}} = 0.138$9 and Spec $0.124$0; prostate reached F1 $0.124$1; rectal tied the best single-model F1 at $0.124$2 while improving AUC to $0.124$3 and Spec to $0.124$4. The same study reports strong global similarity between Prov-GigaPath and H-Optimus1, with CKA up to $0.124$5, but low local neighborhood agreement with $0.124$6-NN overlap below $0.124$7, and identifies Prov-GigaPath together with H-Optimus1 as contributing the most commonly retained features across diseases after pruning (Flannery et al., 11 Dec 2025). This suggests that Prov-GigaPath is particularly effective as a contributor to compact fused representations even when it is not the top standalone model for a given cohort.

6. Limitations, misconceptions, and future directions

The published record identifies several recurring limitations. In LUAD predominant-pattern prediction, only the patch encoder was used, not the full Prov-GigaPath slide-level encoder; class imbalance and limited DHMC sample size, especially papillary with $0.124$8, constrained per-class performance; and hardware, training time, memory footprint, and inference speed were not reported. That study explicitly plans to incorporate the full Prov-GigaPath architecture, including the slide-level encoder, and to extend the framework toward estimating the full distribution of growth patterns with external validation on independent LUAD cohorts (Perez-Herrera et al., 23 Apr 2026). In cell instance segmentation and classification, Prov-GigaPath was evaluated only as a frozen encoder, without stain normalization or fine-tuning, and no CoNIC or CytoDArk0 results were reported because it was not among the five models selected after PanNuke ranking (Vadori et al., 4 Feb 2025). In zero-shot retrieval, GigaPath WSI was processed for only 17 of 23 organs because of computational demands, and subtype-level retrieval remained difficult despite substantial gains over the DenseNet baseline (Alfasly et al., 2024).

Broader analyses raise additional cautionary points. Representational similarity analysis shows that stain normalization decreases slide dependence but does not eliminate it, and Prov-GigaPath retains the highest absolute slide specificity even after Macenko normalization. MGPATH notes persistent sensitivity to domain shift, institution-specific staining and scanner variation, and prompt quality when LLM context is used for textual prompts. The fusion study, for its part, shows that redundancy structure is disease-dependent and that increasingly aggressive pruning can discard large fractions of Prov-GigaPath-derived features at higher correlation thresholds (Mishra et al., 18 Sep 2025, Nguyen et al., 11 Feb 2025, Flannery et al., 11 Dec 2025). A plausible implication is that Prov-GigaPath should not be understood as a universally dominant histopathology backbone; rather, its strengths are most evident when its large-scale pathology pretraining is matched to the task regime—such as fine-tuned ABMIL in LUAD—or when its embeddings are complemented by adapters, structured prompting, or correlation-aware fusion.