Histology images from uniform tumor regions in TCGA Whole Slide Images (TCGA-UT-Internal, TCGA-UT-External)

This repository provides a benchmarking framework for the TCGA histology image dataset originally published on Zenodo. It includes predefined train/validation/test splits and example code for foundation model evaluation.

Task

Classification of 31 different cancer types from tumor histopathological images.

Original Dataset Description

This dataset contains 1,608,060 image patches of hematoxylin & eosin stained histological samples from various human cancers. The data was collected and processed as follows:

• Source: TCGA dataset from 32 solid cancer types (GDC legacy database, downloaded between December 1, 2016, and June 19, 2017)
• Initial data: 9,662 diagnostic slides from 7,951 patients in SVS format
• Annotation: At least three representative tumor regions were selected as polygons by two trained pathologists
• Quality control: 926 slides were removed due to poor staining, low resolution, out-of-focus issues, absence of cancerous regions, or incorrect cancer types
• Final dataset: 8,736 diagnostic slides from 7,175 patients
• Patch extraction: 10 patches at 0.5 μm/pixel resolution (128 x 128 μm) were randomly cropped from each annotated region

Note: Additional resolution levels are available in the original Zenodo dataset. Please refer to the Zenodo repository for the complete dataset.

TCGA Barcode format (TCGA-XX-XXXX) represents patient ID. For details, see the TCGA Barcode documentation.

Updates in This Version

The dataset has been modified and organized for benchmarking purposes:

1. Label Consolidation:

   • Colon Adenocarcinoma (COAD) and Rectum Adenocarcinoma (READ) have been merged due to their histological similarity

2. Structured Splits:

   Internal Split (70:15:15): TCGA-UT-Internal

   • Ensures no patient overlap between train, validation, and test sets
   • Approximate distribution: 70% train, 15% validation, 15% test

   External Split: TCGA-UT-External

   • Separates data based on medical facilities to evaluate cross-institutional generalization
   • No facility overlap between train, validation, and test sets
   • Maintains similar class distributions across splits

Version History

2026-07-01

• Updated setup instructions to use uv.
• Simplified requirements.txt.
• Excluded models pretrained on TCGA from the main benchmark table.
• Added lightweight smoke tests for data loading and evaluation.

Dataset Details

Internal Split: TCGA-UT-Internal

case	train (patches)	valid (patches)	test (patches)	train (patients)	valid (patients)	test (patients)
Adrenocortical_carcinoma	3480	750	750	35	8	8
Bladder_Urothelial_Carcinoma	6990	1500	1500	202	43	44
Brain_Lower_Grade_Glioma	16480	3530	3520	326	70	71
Breast_invasive_carcinoma	16580	3550	3560	513	110	111
Cervical_squamous_cell_carcinoma_and_endocervical_adenocarcinoma	4380	930	960	140	30	31
Cholangiocarcinoma	630	120	150	21	4	5
Colon_Rectum_adenocarcinoma	7020	1510	1500	190	41	41
Esophageal_carcinoma	2360	510	510	78	17	17
Glioblastoma_multiforme	16620	3570	3550	254	54	55
Head_and_Neck_squamous_cell_carcinoma	8250	1770	1770	221	48	48
Kidney_Chromophobe	1710	360	390	57	12	13
Kidney_renal_clear_cell_carcinoma	8160	1740	1750	269	58	58
Kidney_renal_papillary_cell_carcinoma	4750	1020	1020	149	32	33
Liver_hepatocellular_carcinoma	5860	1250	1260	190	41	41
Lung_adenocarcinoma	11520	2470	2470	303	65	66
Lung_squamous_cell_carcinoma	11590	2490	2480	305	66	66
Lymphoid_Neoplasm_Diffuse_Large_B-cell_Lymphoma	570	120	150	19	4	5
Mesothelioma	1470	320	300	42	9	10
Ovarian_serous_cystadenocarcinoma	1740	390	390	58	13	13
Pancreatic_adenocarcinoma	2850	620	620	88	19	19
Pheochromocytoma_and_Paraganglioma	930	210	210	30	7	7
Prostate_adenocarcinoma	6870	1470	1470	212	45	46
Sarcoma	9440	2010	2030	149	32	32
Skin_Cutaneous_Melanoma	7040	1510	1510	226	48	49
Stomach_adenocarcinoma	6770	1450	1450	182	39	39
Testicular_Germ_Cell_Tumors	4210	900	900	92	20	20
Thymoma	2520	540	540	59	13	13
Thyroid_carcinoma	7950	1710	1700	259	56	56
Uterine_Carcinosarcoma	1470	320	330	34	7	8
Uterine_Corpus_Endometrial_Carcinoma	8730	1890	1860	266	57	58
Uveal_Melanoma	1140	240	260	38	8	9
Total	190080	40770	40860	5007	1076	1092

External Split: TCGA-UT-External

case	train (patches)	valid (patches)	test (patches)	train (patients)	valid (patients)	test (patients)
Adrenocortical_carcinoma	4500	390	90	45	5	1
Bladder_Urothelial_Carcinoma	6990	1500	1500	190	50	49
Brain_Lower_Grade_Glioma	16430	3540	3560	332	80	55
Breast_invasive_carcinoma	16560	3570	3560	509	116	109
Cervical_squamous_cell_carcinoma_and_endocervical_adenocarcinoma	4380	930	960	145	31	25
Cholangiocarcinoma	660	150	90	22	5	3
Colon_Rectum_adenocarcinoma	7020	1500	1510	197	39	36
Esophageal_carcinoma	2360	510	510	78	17	17
Glioblastoma_multiforme	16630	3810	3300	244	76	43
Head_and_Neck_squamous_cell_carcinoma	8260	1750	1780	224	51	42
Kidney_Chromophobe	1740	270	450	58	9	15
Kidney_renal_clear_cell_carcinoma	8170	1710	1770	269	57	59
Kidney_renal_papillary_cell_carcinoma	4750	1020	1020	146	34	34
Liver_hepatocellular_carcinoma	5870	1300	1200	189	43	40
Lung_adenocarcinoma	11530	2470	2460	288	77	69
Lung_squamous_cell_carcinoma	11580	2490	2490	296	68	73
Lymphoid_Neoplasm_Diffuse_Large_B-cell_Lymphoma	600	90	150	20	3	5
Mesothelioma	1470	300	320	43	10	8
Ovarian_serous_cystadenocarcinoma	2220	120	180	74	4	6
Pancreatic_adenocarcinoma	2860	600	630	85	20	21
Pheochromocytoma_and_Paraganglioma	1170	90	90	38	3	3
Prostate_adenocarcinoma	6870	1470	1470	226	49	28
Sarcoma	9490	2070	1920	154	28	31
Skin_Cutaneous_Melanoma	7030	1530	1500	233	40	50
Stomach_adenocarcinoma	6990	1330	1350	187	37	36
Testicular_Germ_Cell_Tumors	4600	630	780	96	10	26
Thymoma	2520	540	540	54	18	13
Thyroid_carcinoma	7980	1650	1730	259	54	58
Uterine_Carcinosarcoma	1470	330	320	37	7	5
Uterine_Corpus_Endometrial_Carcinoma	8730	1890	1860	272	48	61
Uveal_Melanoma	1250	120	270	42	4	9
Total	192680	39670	39360	5052	1093	1030

Foundation Model Benchmarking

This benchmark is intended for patch-level feature extractors. Slide-level encoders are outside the scope of this patch dataset.

Models known to include TCGA in pretraining are excluded from the main benchmark table to reduce data contamination risk. Excluded examples include Kaiko/MIDNIGHT, Lunit, Phikon, Phikon-v2, CTransPath, and GenBio-PathFM.

Benchmarked models:

• CONCH
• GigaPath
• UNI
• UNI2
• H-Optimus-0
• H-Optimus-1
• Virchow
• Virchow2
• Hibou
• ImageNet-pretrained ResNet baseline

See licenses/references.txt for model citations.

Benchmark Results

Note: The provided script is a simplified training/evaluation example. The benchmark results below used additional tuning and implementation details.

Internal Split Results

Model	Accuracy (LogReg)	Balanced Accuracy (LogReg)	Accuracy (KNN)	Balanced Accuracy (KNN)	Accuracy (Prototype)	Balanced Accuracy (Prototype)
H-Optimus-1	0.8616	0.8557	0.8164	0.7671	0.7730	0.7579
UNI2	0.8564	0.8501	0.7962	0.7434	0.7546	0.7476
H-Optimus-0	0.8498	0.8399	0.7930	0.7307	0.7492	0.7321
Virchow2	0.8455	0.8351	0.7686	0.6989	0.6671	0.6500
Virchow	0.8223	0.8008	0.7244	0.6262	0.6087	0.5759
Hibou	0.8189	0.7985	0.7433	0.6618	0.6291	0.6034
UNI	0.8144	0.7923	0.7634	0.6897	0.7109	0.6946
GigaPath	0.8161	0.7878	0.7444	0.6676	0.6967	0.6675
CONCH	0.7672	0.7295	0.7028	0.6139	0.6150	0.6097
ResNet	0.6395	0.5581	0.5114	0.3816	0.3154	0.2973

External Split Results

Model	Accuracy (LogReg)	Balanced Accuracy (LogReg)	Accuracy (KNN)	Balanced Accuracy (KNN)	Accuracy (Prototype)	Balanced Accuracy (Prototype)
H-Optimus-1	0.8080	0.7450	0.7700	0.6955	0.7572	0.7363
UNI2	0.7648	0.7262	0.7210	0.6498	0.7018	0.6839
H-Optimus-0	0.7845	0.7213	0.7209	0.6579	0.7106	0.6842
Virchow2	0.7744	0.6919	0.7221	0.6544	0.6482	0.6331
UNI	0.7373	0.6581	0.6668	0.5887	0.6612	0.6232
Virchow	0.7274	0.6490	0.6464	0.5541	0.5847	0.5636
GigaPath	0.7246	0.6379	0.6426	0.5495	0.6361	0.5960
Hibou	0.6696	0.6161	0.5155	0.4436	0.4911	0.4765
CONCH	0.6991	0.5975	0.6626	0.5735	0.5954	0.5905
ResNet	0.4967	0.3929	0.3960	0.2871	0.2657	0.2392

Getting Started

1. Clone this repository:

git clone https://huggingface.co/datasets/dakomura/tcga-ut
cd tcga-ut

2. Install dependencies with uv:

uv venv --python 3.10
source .venv/bin/activate
uv pip install -r requirements.txt

If you do not use uv, pip install -r requirements.txt also works in a compatible Python environment.

3. Download Git LFS files and log in to Hugging Face:

git lfs pull
hf auth login

The default example uses bioptimus/H-optimus-0, which is a gated model. Your Hugging Face account must have access to the dataset and the selected model.

4. Run the experiment:

python extract_train.py

Optional quick checks before a full run:

python scripts/smoke_test_data_loading.py --shard data/dataset_internal_test_part000.tar
python scripts/smoke_test_eval.py

Verified Setup

We verified uv pip install -r requirements.txt with Python 3.10 on Linux. The data-loading smoke test loaded one materialized WebDataset shard, and the eval-only smoke test ran logreg, knn, and proto on synthetic features. Full feature extraction requires the full Git LFS dataset, a selected model with access permission, and a CUDA-capable environment for the default config.yaml.

Troubleshooting

• Git LFS pointer files: If CSV or tar files contain version https://git-lfs.github.com/spec/v1, run git lfs pull.
• Gated model access: A 403 GatedRepoError means your Hugging Face account does not have access to the selected model.
• No CUDA device: Set device: "cpu" in config.yaml for small checks. Full feature extraction on CPU will be slow.
• SPAMS installation: spams is not required for extract_train.py. If you need SPAMS for separate stain-normalization workflows, try pip install spams-bin.

Data Loading Example

The dataset uses WebDataset format for efficient loading. Here's an example from extract_train.py:

patterns = {
    'train': [os.path.join(work_dir, f"data/dataset_{split}_train_part{str(i).zfill(3)}.tar") for i in range(39)],
    'valid': [os.path.join(work_dir, f"data/dataset_{split}_valid_part{str(i).zfill(3)}.tar") for i in range(file_range)],
    'test': [os.path.join(work_dir, f"data/dataset_{split}_test_part{str(i).zfill(3)}.tar") for i in range(file_range)],
}
dataset = wds.WebDataset(patterns[mode], shardshuffle=False) \
    .shuffle(buffer_size, seed=42) \
    .decode("pil").to_tuple("jpg", "json") \
    .map_tuple(func_transform, lambda x: encode_labels([x["label"]], label_encoder))

Configuration and Usage

1. Configure your experiment in config.yaml:

model_name: "h_optimus"    # Model selection: "h_optimus", etc.
split_type: "internal"     # Split type: "internal" or "external"
device: "cuda"            # Computation device: "cuda" or "cpu"
eval_name: "logreg"       # Evaluation method: "logreg", "knn", or "proto"
feature_exist: True       # Skip feature extraction if features already exist
max_iter: 1000           # Maximum iterations for training
cost: 0.0001             # Cost parameter for logistic regression

Configuration parameters:

• model_name: Foundation model to use for feature extraction
• split_type: Dataset split strategy
• eval_name: Methods of evaluation (logreg, knn, proto)
• device: Computation device (GPU/CPU)
• feature_exist: Skip feature extraction if True and features are already available
• max_iter: Maximum training iterations for logistic regression
• cost: Regularization parameter for logistic regression
• k: Number of Nearest Neighbors in KNN

2. Define models and transforms in extract_train.py:

def get_model_transform(model_name):
    # Add your model and transform definitions here
    pass

3. Run the experiment:

python extract_train.py

This will:

• Extract features using the specified foundation model
• Save features to H5 files
• Perform linear probing, KNN, and prototype classification
• Output accuracy and balanced accuracy metrics

License

This work is licensed under a Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International (CC-BY-NC-SA 4.0).

• For non-commercial use: Please use the dataset under CC-BY-NC-SA
• For commercial use: Please contact us at ishum-prm@m.u-tokyo.ac.jp

Citation

If you use this dataset, please cite the original paper:

@article{komura2022universal,
  title={Universal encoding of pan-cancer histology by deep texture representations},
  author={Komura, D., Kawabe, A., Fukuta, K., Sano, K., Umezaki, T., Koda, H., Suzuki, R., Tominaga, K., Ochi, M., Konishi, H., Masakado, F., Saito, N., Sato, Y., Onoyama, T., Nishida, S., Furuya, G., Katoh, H., Yamashita, H., Kakimi, K., Seto, Y., Ushiku, T., Fukayama, M., Ishikawa, S.},
  journal={Cell Reports},
  volume={38},
  pages={110424},
  year={2022},
  doi={10.1016/j.celrep.2022.110424}
}