在医学诊断领域，准确、快速的病理分析对于治疗计划和患者预后至关重要。随着人工智能技术的发展，计算病理学领域迎来了革命性的进步，其中预训练模型在提高诊断准确性、预后评估和治疗反应预测方面发挥着关键作用。关注微信公众号：欣欣影像科研懒人包
预训练模型通过在大规模数据集上学习特征表示，能够显著减少对标注数据的依赖，降低成本并加速新模型的开发。这些模型不仅能够识别已知疾病，还能够在罕见和复杂病例中提供诊断支持。
近年来，病理学中的预训练模型主要可以分为切片级（Patch-level）模型和全切片级（Slide-level）模型。切片级模型通过将整张病理玻片切割成多个小的子切片进行分析，重点关注局部区域的病理特征。而全切片级模型则直接对整张切片进行处理，利用全局信息进行更全面的分析。这两类模型各有特点，切片级模型在局部特征识别上具有较强优势，而全切片级模型则能整合整个切片的信息。
接下来的表格整理了近年来最先进的病理组学特征提取模型以及相关开源的论文及模型权重地址，以供相关研究人员进行改进或作下游微调任务。关注微信公众号：欣欣影像科研懒人包
| 切片级特征提取器 （Patch-level models）关注微信公众号：欣欣影像科研懒人包 | ||||||||
| 模型名称 | 作者/团队 | 权重 | 释放日期 | 基模型 | 全切片数量（WSIs） | 切片数量（Tiles） | 架构 | 数据集 |
|---|---|---|---|---|---|---|---|---|
| CTransPath | Sichuan University / Tencent AI Lab | ✔ | Dec 2021* | SRCL | 32K | 16M | Swin-Transformer | TCGA, PAIP |
| RetCCL | Sichuan University / Tencent AI Lab | ✔ | Dec 2021* | CCL | 32K | 16M | ResNet-50 | TCGA, PAIP |
| REMEDIS | Google Research | ✔ | May 2022* | SimCLR/BiT | 29K | 50M | ResNet-50 | TCGA |
| HIPT | Mahmood Lab | ✔ | Jun 2022* | DINOv1 | 11K | 100M | ViT-S | TCGA |
| Lunit-DINO | Lunit | ✔ | Dec 2022* | DINOv1 | 21K | ViT-S | TCGA | |
| Lunit-{BT,MoCoV2,SwAV} | Lunit | ✔ | Dec 2022* | {BT,MoCoV2,SwAV} | 21K | ResNet-50 | TCGA | |
| Phikon | Owkin | ✔ | Jul 2023* | iBOT | 6.1K | 43M | ViT-B | TCGA |
| CONCH | Mahmood Lab | ✔ | Jul 2023* | iBOT & vision-language pretraining | 21K | 16M | ViT-B | proprietary |
| UNI | Mahmood Lab | ✔ | Aug 2023* | DINOv2 | 100K | 100M | ViT-L | proprietary (Mass-100K) |
| Virchow | Paige / Microsoft | ✔ | Sep 2023* | DINOv2 | 1.5M | ViT-H | proprietary (from MSKCC) | |
| Campanella et al. (MAE) | Thomas Fuchs Lab | ❌ | Oct 2023* | MAE | 420K | 3.3B | ViT-L | proprietary (MSHS) |
| Campanella et al. (DINO) | Thomas Fuchs Lab | ❌ | Oct 2023* | DINOv1 | 420K | 3.3B | ViT-L | proprietary (MSHS) |
| Path Foundation | | ✔ | Oct 2023* | SimCLR, MSN | 6K | 60M | ViT-S | TCGA |
| PathoDuet | Shanghai Jiao Tong University | ✔ | Dec 2023* | inspired by MoCoV3 | 11K | 13M | ViT-B | TCGA |
| RudolfV | Aignostics | ❌ | Jan 2024* | DINOv2 | 100K | 750M | ViT-L | proprietary (from EU & US), TCGA |
| kaiko | kaiko.ai | ✔ | Mar 2024* | DINOv2 | 29K | 260M** | ViT-L | TCGA |
| PLUTO | PathAI | ❌ | May 2024* | DINOv2 (+ MAE and Fourier loss) | 160K | 200M | FlexiViT-S | proprietary (PathAI) |
| BEPH | Shanghai Jiao Tong University | ✔ | May 2024* | BEiTv2 | 12K | 12M | ViT-B | TCGA |
| Prov-GigaPath | Microsoft / Providence | ✔ | May 2024* | DINOv2 | 170K | 1.4B | ViT | proprietary (Providence) |
| Hibou-B | HistAI | ✔ | Jun 2024* | DINOv2 | 1.1M | 510M | ViT-B | proprietary |
| Hibou-L | HistAI | ✔ | Jun 2024* | DINOv2 | 1.1M | 1.2B | ViT-L | proprietary |
| H-optimus-0 | Bioptimus | ✔ | Jul 2024* | DINOv2/iBOT | 500K (across 4,000 clinics) | >100M | ViT-G with 4 registers | proprietary |
| mSTAR | Smart Lab | ❌ | Jul 2024* | mSTAR (multimodal) | 10K | ViT-L | TCGA | |
| Virchow 2 | Paige / Microsoft | ✔ | Aug 2024* | DINOv2 (+ ECT and KDE) | 3.1M | 2B | ViT-H with 4 registers | proprietary (from MSKCC and international sites) |
| Virchow 2G | Paige / Microsoft | ❌ | Aug 2024* | DINOv2 (+ ECT and KDE) | 3.1M | 2B | ViT-G with 8 registers | proprietary (from MSKCC and international sites) |
| Phikon-v2 | Owkin | ✔ | Sep 2024* | DINOv2 | 58.4K | 456M | ViT-L | PANCAN-XL (TCGA, CPTAC, GTEx, proprietary) |
| 全切片级特征提取模型 Slide-level / patient-level models 关注微信公众号：欣欣影像科研懒人包 | ||||||
| 模型名称 | 作者/团队 | 释放日期 | 基模型 | 全切片数量（WSIs） | 架构 | 数据集 |
|---|---|---|---|---|---|---|
| GigaSSL | CBIO | Dec 2022* | SimCLR | 12K | ResNet-18 | TCGA |
| PRISM | Paige / Microsoft | May 2024* | contrastive (with language) | 590K (190K text reports) | Perceiver + BioGPT | proprietary |
| Prov-GigaPath | Microsoft / Providence | May 2024* | DINOv2 | 170K | LongNet | proprietary (Providence) |
| MADELEINE | Mahmood Lab | Aug 2024* | contrastive (InfoNCE & OT) | 16K | multi-head attention MIL | ACROBAT, BWH Kidney (proprietary) |
| CHIEF | Yu Lab | Sep 2024* | ||||
| 模型名称 | 论文链接 关注微信公众号：欣欣影像科研懒人包 |
| CTransPath | https://www.sciencedirect.com/science/article/abs/pii/S1361841522002043 |
| RetCCL | https://www.sciencedirect.com/science/article/abs/pii/S1361841522002043 |
| REMEDIS | https://www.nature.com/articles/s41551-023-01049-7 |
| HIPT | https://ieeexplore.ieee.org/document/9880275 |
| Lunit-DINO | https://arxiv.org/abs/2212.04690 |
| Lunit-{BT,MoCoV2,SwAV} | https://arxiv.org/abs/2212.04690 |
| Phikon | https://www.medrxiv.org/content/10.1101/2023.07.21.23292757v2 |
| CONCH | https://www.nature.com/articles/s41591-024-02856-4 |
| UNI | https://www.nature.com/articles/s41591-024-02857-3 |
| Virchow | https://www.nature.com/articles/s41591-024-03141-0 |
| Campanellaet al.(MAE) | https://arxiv.org/abs/2310.07033 |
| Campanellaet al.(DINO) | https://arxiv.org/abs/2310.07033 |
| Path Foundation | https://arxiv.org/abs/2310.13259 |
| PathoDuet | https://arxiv.org/abs/2312.09894 |
| RudolfV | https://arxiv.org/abs/2401.04079 |
| kaiko | https://arxiv.org/abs/2404.15217 |
| PLUTO | https://arxiv.org/abs/2405.07905 |
| BEPH | https://www.biorxiv.org/content/10.1101/2024.05.16.594499 |
| Prov-GigaPath | https://www.nature.com/articles/s41586-024-07441-w |
| Hibou-B | https://arxiv.org/abs/2406.05074 |
| Hibou-L | https://arxiv.org/abs/2406.05074 |
| H-optimus-0 | https://www.bioptimus.com/news/bioptimus-launches-h-optimus-0-the-worlds-largest-open-source-ai-foundation-model-for-pathology |
| mSTAR | https://arxiv.org/abs/2407.15362 |
| Virchow 2 | https://arxiv.org/abs/2408.00738 |
| Virchow 2G | https://arxiv.org/abs/2408.00738 |
| Phikon-v2 | https://arxiv.org/abs/2409.09173 |
| 模型名称 | 论文链接 |
| GigaSSL | https://arxiv.org/abs/2212.03273 |
| PRISM | https://arxiv.org/abs/2405.10254 |
| Prov-GigaPath | https://www.nature.com/articles/s41586-024-07441-w |
| MADELEINE | https://arxiv.org/abs/2408.02859 |
| CHIEF | https://www.nature.com/articles/s41586-024-07894-z |
| 模型名称 | 模型权重链接 关注微信公众号：欣欣影像科研懒人包 |
| CTransPath | https://github.com/Xiyue-Wang/TransPath |
| RetCCL | https://github.com/Xiyue-Wang/RetCCL |
| REMEDIS | https://github.com/google-research/medical-ai-research-foundations |
| HIPT | https://github.com/mahmoodlab/HIPT |
| Lunit-DINO | https://github.com/lunit-io/benchmark-ssl-pathology |
| Lunit-{BT,MoCoV2,SwAV} | https://github.com/lunit-io/benchmark-ssl-pathology |
| Phikon | https://github.com/owkin/HistoSSLscaling https://huggingface.co/owkin/phikon |
| CONCH | https://github.com/mahmoodlab/CONCH https://huggingface.co/MahmoodLab/CONCH |
| UNI | https://github.com/mahmoodlab/UNI https://huggingface.co/MahmoodLab/UNI |
| Virchow | https://huggingface.co/paige-ai/Virchow |
| Campanella et al. (MAE) | X |
| Campanella et al. (DINO) | X |
| Path Foundation | https://github.com/Google-Health/imaging-research/tree/master/path-foundation |
| PathoDuet | https://github.com/openmedlab/PathoDuet |
| RudolfV | X |
| kaiko | X |
| PLUTO | X |
| BEPH | https://github.com/Zhcyoung/BEPH |
| Prov-GigaPath | https://github.com/prov-gigapath/prov-gigapath https://huggingface.co/prov-gigapath/prov-gigapath |
| Hibou-B | https://github.com/HistAI/hibou https://huggingface.co/histai/hibou-b |
| Hibou-L | https://github.com/HistAI/hibou https://huggingface.co/histai/hibou-L |
| H-optimus-0 | https://github.com/bioptimus/releases/tree/main/models/h-optimus/v0 |
| mSTAR | X |
| Virchow 2 | https://huggingface.co/paige-ai/Virchow2 |
| Virchow 2G | X |
| Phikon-v2 | https://huggingface.co/owkin/phikon-v2 |
| 模型名称 | 模型链接 |
| GigaSSL | https://github.com/trislaz/gigassl |
| PRISM | https://huggingface.co/paige-ai/Prism |
| Prov-GigaPath | https://github.com/prov-gigapath/prov-gigapath |
| MADELEINE | https://github.com/mahmoodlab/MADELEINE |
| CHIEF | https://github.com/hms-dbmi/CHIEF |
声明：本文整理于Pathology Feature Extractors and Foundation Models(https://github.com/georg-wolflein/pathology-foundation-models)，原创声明仅代表对原文内容的整理和输出，版权归原作者所有，如有侵权请联系删除，感谢。关注微信公众号：欣欣影像科研懒人包
扫一扫
1219
被折叠的 条评论
为什么被折叠?
到【灌水乐园】发言
