|
Download README.md from Pathstudy/omni_pathology: direct link, hf CLI and curl.
- Browser
- Download file 3.75 kB
-
https://huggingface.co/datasets/Pathstudy/omni_pathology/resolve/main/README.md
- Command line
-
hf download hf://datasets/Pathstudy/omni_pathology/README.md
-
curl -L -H "Authorization: Bearer $HF_TOKEN" -o README.md https://huggingface.co/datasets/Pathstudy/omni_pathology/resolve/main/README.md
3.75 kB
| license: mit | |
| # path_example — 病理多模态数据 ~1% 子样本 | |
| 来源:Misha 集群 `/gpfs/radev/pi/ying_rex/tl688` 与 `/gpfs/radev/project/zhao/tl688`。 | |
| 生成脚本:`_scripts/subsample.py`(固定随机种子 0,每个模态取约 1%,极小集合设最低条数下限)。 | |
| 每个数据集的实际抽样条数 / 对象结构记录在 `_manifest/*.json`,运行日志在 `_logs/`。 | |
| 目录布局与原始数据保持一致,文件名不变,可直接替换路径前缀使用。 | |
| ## 抽样规则 | |
| | 类型 | 规则 | | |
| |---|---| | |
| | 图像 tile / patch | 随机取 1%(ORION 按切片分层,每张切片 1%) | | |
| | 配对模态(IF、核 mask、mIHC、ROSIE 预测、spot 图像、特征) | 与主模态使用**同一索引**,保持配对 | | |
| | adata / h5ad | 每个样本随机 1% spot(最少 5–20 个) | | |
| | 大型 json / csv(PathGen、QA、标签表) | 随机 1% 行/条目(最少 20 条) | | |
| | 小型索引/元数据(slide_dataframe、HEST csv、allwsi.txt、manifest、antibody 表、README) | 整体复制 | | |
| | 单张荧光大图 | 取中心 10%×10%(面积 1%)裁剪 | | |
| | 模型权重、代码、实验输出目录 | 不包含 | | |
| ## 目录 | |
| - `ORION/ORIONCRC_dataset_tile_20x/` — he/ if/ nuclei/ 配对 tile;train/val/test_dataframe.csv 已过滤;csv_nuclei_pos/ 仅保留落在所选 tile 内的核(按核 mask label 匹配) | |
| - `HEMIT/` — he/{train,val,test}/{input,label} 配对 tif;HEMIT_nuclei_analysis 对应 csv/mask;`ROSIE/rosie_pred_hemit_subset.npy` 为 ROSIE 对所选 test patch 的预测(原 npy 第 0 维按 test_data.csv 行序,索引见 `rosie_pred_hemit_subset_index.csv`) | |
| - `HEST/` — `visium_humandisease_adata.pickle`:421 个 Visium 样本的 AnnData 列表,每样本随机 1% spot(共 10,233 spot);`xenium_humandisease_adata.pickle`:59 个样本,每样本 1%;`UNI/`、`prov-gigapath-conchinside/` 中 UNIv2 / GigaPath 特征列表与 adata 逐样本同索引(421/421 对齐)。注意:每个 AnnData 的 `uns['spatial']` 保留了原样本的 downscaled_fullres 缩略图(每张约 644×1000×3),这是子集 2.2G 体积的主要来源 | |
| - `DLPFC_visium/` — 12 样本(151507–151676,共 47,681 spot)各模型各 patch size 的 spot 特征 158 个 pkl(同一样本共用同一 spot 索引,见 `_manifest/dlpfc.json`);`GPFM/visium_image_list.pkl`(482×3×224×224)、`visium_gene_expression_list.pkl`(482×10000)、`hsit2pix_out_result.npy`(482×3×224×224)按全局 spot 索引(样本按 id 升序拼接,见 `global_spot_index.csv`)同步抽取;`visium_fluo_image_crop_1pct_area.npy` 为荧光图中心 727×727 裁剪。`GPFM/visium_gpfm_imagefeature_200.pickle` 实际为 421 个 batch × 1024 行(与 HEST visium 样本数一致,非 DLPFC),每 batch 独立取 1% | |
| - `IDC/UNI/` — IDC_adata + IDC_image(同索引) | |
| - `PathGen-1.6M/` — PathGen-1.6M.json(16,000 条)、PathGen-Instruct.json 1% 条目;figure_caption_summary 30 张 patch + 对应 caption 行;各癌种(idc / gliomas / saromas / skin / brain)train json 1% 条目 + 其引用的 patch;test patch 目录 1%;stimagedata h5ad 每个 1% spot,`_new.pkl`(分块的 PIL spot 图像列表)按 obs 顺序同索引;gliomas / saromas 的 test h5ad 与对应 test patch 目录按序号对齐;train/valid/test_idc.h5ad 1%;slide_report 每 set 1 个 patch + desc + 对应预测;QA/CoT/PathMMU 输出 csv 1% 行 | |
| - `foundation_features/` — GPFM / UNI 的 COAD slide embedding(同索引);GPFM dataset_csv 与 splits* 每个 csv 1% 行(保留完整目录结构) | |
| - `STPath/` — example_data/INT2.h5ad 1% spot、标注 json(整体复制) | |
| - `path_report_database/` — caption csv 1% 行 | |