README.md

metadata

license: cc-by-nc-4.0
language:
  - en
library_name: pytorch
pipeline_tag: image-classification
metrics:
  - accuracy
  - auc
  - f1

DigeApplication

DigeApplication is an open-source tool-model zoo for downstream gastrointestinal computational pathology tasks, providing unified configs and inference utilities for rapid deployment and reproducible evaluation.

• Foundation model: Digepath (ViT-L/16, DINOv2 pretraining)
• Digepath paper (arXiv): https://arxiv.org/abs/2505.21928

What is DigeApplication?

DigeApplication packages a collection of task-specific tool-models built on top of gastrointestinal pathology foundation model (Digepath). It aims to provide:

• Model Zoo: standardized, ready-to-use downstream tool-models
• Unified Configs: each tool is described by a lightweight config.json
• Unified Inference: consistent API to load and run a tool-model

Repository Structure

DigeApplication/
  ├── Model_Zoo/
  │     ├── <tool_name_1>/
  │     │     ├── config.json
  │     │     ├── model.safetensors
  │     └── <tool_name_2>/
  ├── Model_Factory/
  │     ├── digepath_model_factory.py
  │     └── digepath_modeling_wsi.py
  │     └── digepath_modeling_roi.py
  ├── DigeApplication.py
  └── README.md

Usage Example

from Model_Factory.digepath_model_factory import DigepathModelFactory
import os
import torch

model_zoo_dir = "./DigeApplication/Model_Zoo/"
device = "cuda:0"
dummy_num_patches = 100
model_dirs = [
    os.path.join(model_zoo_dir, d)
    for d in os.listdir(model_zoo_dir)
    if os.path.isdir(os.path.join(model_zoo_dir, d))
]
for digemodel_dir in model_dirs:
    print(f"Loading model from {digemodel_dir}...")
    model_factory = DigepathModelFactory(digemodel_dir, device=device)
    if model_factory.data_type == "ROI":
        dummy_input = torch.randn(1, 1024, device=device)
    elif model_factory.data_type == "WSI":
        dummy_input = torch.randn(1, dummy_num_patches, 1024, device=device)
    else:
        raise ValueError(f"Unknown data_type: {model_factory.data_type}")
    output = model_factory.infer(dummy_input)
    print(f"Model output: {output}\n")

Supported Tasks (Model Zoo)

Each tool-model in DigeApplication is described by a config containing fields such as: tool_name, data_type, task_type, organ_domain, num_classes

Tool / Model Name	Task Type	Data Type	Organ Domain	Classes	License
BOW_CAMEL_2CLS_ROI	CLS	ROI	BOW	2	cc_by_nc_4.0
BOW_CENPL_5CLS_ROI	CLS	WSI	BOW	5	cc_by_nc_4.0
BOW_COAD-MUC16_2CLS_WSI	CLS	WSI	BOW	2	cc_by_nc_4.0
BOW_COAD-TTN_2CLS_WSI	CLS	WSI	BOW	2	cc_by_nc_4.0
BOW_CRC100K_9CLS_ROI	CLS	ROI	BOW	9	cc_by_nc_4.0
BOW_READ-NRAS_2CLS_WSI	CLS	WSI	BOW	2	cc_by_nc_4.0
BOW_READ-TTN_2CLS_WSI	CLS	WSI	BOW	2	cc_by_nc_4.0
BOW_TNM_4CLS_WSI	CLS	WSI	BOW	4	cc_by_nc_4.0
DIGE_HER2_2CLS_WSI	CLS	WSI	DIGE	2	cc_by_nc_4.0
DIGE_IMPCRS_3CLS_ROI	CLS	WSI	DIGE	3	cc_by_nc_4.0
DIGE_MSI_2CLS_WSI	CLS	WSI	DIGE	2	cc_by_nc_4.0
DIGE_P53_2CLS_WSI	CLS	WSI	DIGE	2	cc_by_nc_4.0
DIGE_PDL1_2CLS_WSI	CLS	WSI	DIGE	2	cc_by_nc_4.0
DIGE_POORAS_2CLS_WSI	CLS	WSI	DIGE	2	cc_by_nc_4.0
DIGE_TUM_LHN_3CLS_WSI	CLS	WSI	DIGE	3	cc_by_nc_4.0
DIGE_UNITOPATHO_6CLS_ROI	CLS	ROI	DIGE	6	cc_by_nc_4.0
ESO_ADESQUA_2CLS_WSI	CLS	WSI	ESO	2	cc_by_nc_4.0
ESO_EPILES_2CLS_WSI	CLS	WSI	ESO	2	cc_by_nc_4.0
ESO_EPILES_4CLS_WSI	CLS	WSI	ESO	4	cc_by_nc_4.0
ESO_KERATIN_2CLS_WSI	CLS	WSI	ESO	2	cc_by_nc_4.0
GAST_ACT_3CLS_WSI	CLS	WSI	GAST	3	cc_by_nc_4.0
GAST_ATR_2CLS_WSI	CLS	WSI	GAST	2	cc_by_nc_4.0
GAST_GASTRITIS_3CLS_WSI	CLS	WSI	GAST	3	cc_by_nc_4.0
GAST_IM_4CLS_WSI	CLS	WSI	GAST	4	cc_by_nc_4.0
GAST_STLC_11CLS_ROI	CLS	ROI	GAST	11	cc_by_nc_4.0
GAST_TNM_4CLS_WSI	CLS	WSI	GAST	4	cc_by_nc_4.0
GAST_XANSRC_2CLS_WSI	CLS	WSI	GAST	2	cc_by_nc_4.0

Citation

If DigeApplication is helpful to you, please cite our work.

@article{zhu2025subspecialty,
  title={Subspecialty-specific foundation model for intelligent gastrointestinal pathology},
  author={Zhu, Lianghui and Ling, Xitong and Ouyang, Minxi and Liu, Xiaoping and Guan, Tian and Fu, Mingxi and Cheng, Zhiqiang and Fu, Fanglei and Zeng, Maomao and Liu, Liming and others},
  journal={arXiv preprint arXiv:2505.21928},
  year={2025}
}