From Learning the Basics to Conducting Reproducible Medical AI Research
一个面向医学人工智能初学者与科研人员的系统化学习与科研指南
Medical-AI-Guide 是一个面向医学人工智能学习者与科研人员的开源指南。
本项目最初希望帮助医学人工智能初学者快速建立完整的技术栈；随着医学 AI 从传统深度学习逐渐发展到 Foundation Models、Multimodal Learning、Large Language Models、Vision-Language Models、Diffusion Models 与 Generative AI，本项目也将内容逐步扩展为更加完整的医学 AI 科研工作流。
我们希望你在这里获得的不只是“如何运行代码”，更是：
如何发现问题 → 理解数据 → 设计方法 → 实现模型 → 正确评估 → 撰写论文 → 发布代码 → 做出可复现的医学 AI 研究。
欢迎 ⭐ Star、💡 提交 Issue、🔀 Pull Request 与分享本项目。
如果你是第一次接触医学人工智能，推荐按照下面的路线逐步学习：
Mathematics & Programming
↓
Machine Learning
↓
Deep Learning
↓
Medical Imaging
↓
Medical AI Research
↓
┌────────┼──────────┐
↓ ↓ ↓
CV Generative Multimodal
AI / Diff. AI
↓ ↓ ↓
Foundation Models / LLMs / VLMs
↓
Medical AI Research
↓
Evaluation & Reproducibility
↓
Paper Writing
↓
Open-source / Deployment
医学 AI 科研的第一步不是写代码，而是正确理解问题与建立研究视野。
- Google Scholar — 学术论文搜索与引用追踪
- Semantic Scholar — AI 驱动的论文搜索与推荐
- arXiv — 预印本论文平台
- OpenReview — 会议论文、评审与讨论
- PubMed — 生物医学论文检索
- AMiner — 学术资源搜索与学术网络分析
- Connected Papers — 论文关系图谱
- DeepWiki — AI-assisted GitHub Repository 理解
- Medical Image Analysis
- Medical Image Segmentation
- Medical Image Classification
- Object Detection & Localization
- Image Registration
- Image Reconstruction
- Image Translation
- Generative AI
- Diffusion Models
- Foundation Models
- Large Language Models (LLMs)
- Vision-Language Models (VLMs)
- Multimodal Learning
- Medical Vision-Language Learning
- Medical Image Generation
- Virtual Staining
- Computational Pathology
- Radiology AI
- Clinical Decision Support
推荐不要只关注“模型用了什么结构”，而是重点回答：
What is the problem?
↓
Why does it matter?
↓
What are the limitations of previous methods?
↓
What is the key idea?
↓
Why should the proposed method work?
↓
How is it evaluated?
↓
Does the evidence really support the claim?
一个可靠的科研环境应该同时考虑：
可用性 + 依赖管理 + 可复现性 + GPU + 远程开发
在现代 Python 项目中，可以优先考虑
pyproject.toml+ lockfile 的项目管理方式，而不是依赖一份长期漂移的requirements.txt。
- SSH
- tmux
- VS Code Remote SSH
- Jupyter
- Docker
- GPU server / workstation
- Cloud GPU
- PyTorch
- torchvision
- PyTorch Lightning
- MONAI — 医学影像 AI 开发框架
- NiBabel — NIfTI 等神经影像数据处理
- SimpleITK — 医学图像读取、处理与转换
重点掌握：
- CNN
- Vision Transformer
- Attention
- Object Detection
- Semantic Segmentation
- Instance Segmentation
- Image Registration
- Representation Learning
- Contrastive Learning
- Self-supervised Learning
如今的医学 AI 已经不应只围绕传统 CNN 分类与分割展开。
- GAN
- VAE
- Normalizing Flow
- Diffusion Models
- Score-based Models
- Flow Matching
- Rectified Flow
- Bridge / Schrödinger Bridge Models
推荐重点理解：
Forward Process
↓
Noise / Corruption
↓
Denoising Network
↓
Reverse Process
↓
Generated / Translated Image
进一步学习：
- DDPM
- DDIM
- Latent Diffusion
- Conditional Diffusion
- Diffusion Transformer
- Image-to-Image Diffusion
- Video Diffusion
- Diffusion for Medical Imaging
- Diffusion-based Reconstruction
- Diffusion-based Translation
- Diffusion-based Image Enhancement
医学 AI 正逐渐从“为一个任务训练一个模型”转向：
Pretrain → Adapt → Evaluate → Deploy
重点关注：
- Vision Foundation Models
- Vision Transformers
- Self-supervised Pretraining
- Large-scale Pretrained Encoders
- Medical Foundation Models
- Domain Adaptation
- Parameter-efficient Fine-tuning
重点掌握：
- Prompt Engineering
- Structured Output
- Function Calling
- Tool Use
- Retrieval-Augmented Generation
- Long-context Modeling
- LLM-based Information Extraction
- LLM-based Evaluation
- Agentic Workflows
重点关注：
- Image-Text Alignment
- Contrastive Learning
- Visual Question Answering
- Image Captioning
- Medical VQA
- Radiology Report Understanding
- Multimodal Reasoning
- Vision-Language Agents
- Chest X-ray
- Bone Suppression
- Image Enhancement
- Reconstruction
- Disease Classification
- Abnormality Detection
- 2D / 3D Image Processing
- Segmentation
- Registration
- Reconstruction
- Volumetric Analysis
- Image Classification
- Segmentation
- Video Analysis
- Reconstruction
- Whole Slide Imaging
- Computational Pathology
- Cell Detection
- Tissue Classification
- Virtual Staining
- Spatial Analysis
Classification
Detection
Segmentation
Registration
Reconstruction
Generation
Translation
Report Generation
Visual Question Answering
Multimodal Reasoning
Clinical Prediction
医学 AI 最大的挑战之一往往并不是模型，而是数据。
推荐系统学习：
- Dataset Construction
- Data Cleaning
- Data Annotation
- Data Quality Control
- Patient-level Splitting
- Train / Validation / Test
- Cross-validation
- Data Leakage
- Class Imbalance
- External Validation
- Domain Shift
- Multi-center Validation
- Public Dataset Reproducibility
一个医学 AI 模型“效果更好”，并不等于实验结果可信。
除了常见指标之外，应该重点学习：
- Accuracy
- Precision
- Recall
- F1-score
- AUROC
- AUPRC
- Sensitivity
- Specificity
- Calibration
- Dice
- IoU
- HD95
- ASSD
- PSNR
- SSIM
- LPIPS
- FID
- KID
- NIQE
- BRISQUE
同时注意：
- Statistical Significance
- Confidence Intervals
- Bootstrap
- Wilcoxon Signed-rank Test
- Paired / Unpaired Statistical Tests
- Multiple Comparisons
- Effect Size
机器指标并不是终点。
进一步学习：
- Reader Study
- Human Evaluation
- Radiologist / Pathologist Assessment
- Inter-rater Agreement
- Cohen's Kappa
- Intraclass Correlation
- Clinical Utility
- External Validation
优秀的医学 AI 论文不只是“模型更复杂”。
应该形成完整的实验链：
Research Question
↓
Hypothesis
↓
Method
↓
Baseline
↓
Ablation Study
↓
Quantitative Evaluation
↓
Qualitative Evaluation
↓
Statistical Analysis
↓
External / Clinical Validation
↓
Conclusion
重点关注：
- Strong Baselines
- Fair Comparison
- Ablation Study
- Sensitivity Analysis
- Robustness Analysis
- Generalization
- Computational Cost
- Inference Efficiency
- Failure Cases
- Reproducibility
如今的医学 AI 科研越来越强调：
Can someone else reproduce your result?
建议所有项目包含：
project/
├── configs/
├── datasets/
├── models/
├── scripts/
├── src/
├── tests/
├── notebooks/
├── README.md
├── pyproject.toml
└── LICENSE
同时记录：
- Python version
- PyTorch version
- CUDA version
- GPU
- Random seed
- Dataset version
- Model checkpoint
- Training configuration
- Hyperparameters
- Evaluation protocol
- Git
- GitHub
- Git LFS
- Docker
- uv
- DVC
- Weights & Biases
- MLflow
科研最终需要通过论文、报告与公开代码完成交流。
推荐掌握：
- Matplotlib
- Plotly
- Seaborn
- Illustrator
- Inkscape
- PowerPoint
- draw.io
- Mermaid
重点不是“把图做得漂亮”，而是：
让读者在最短时间内理解你的方法与实验。
AI 已经逐渐成为科研工作流的一部分，但应该把它作为：
Research Assistant，而不是 Research Replacement。
可以用于：
- Literature Search
- Paper Summarization
- Code Understanding
- Debugging
- Documentation
- Information Extraction
- Data Processing
- Figure Drafting
- Writing Assistance
- Translation
- Brainstorming
同时必须保持：
- Source Verification
- Experimental Verification
- Citation Verification
- Code Verification
- Data Privacy
- Clinical Safety Awareness
AI 生成的内容永远不能替代研究者对实验结果与科学结论的最终判断。
| Category | Tools |
|---|---|
| 📝 Writing | LaTeX · Overleaf · Markdown |
| 🔬 Research | Google Scholar · PubMed · arXiv · OpenReview |
| 🐍 Python | Python · uv · Conda · pip |
| 🧠 Deep Learning | PyTorch · MONAI |
| 👁️ Vision | OpenCV · torchvision · timm |
| 🤗 Foundation Models | Hugging Face · Transformers · Diffusers |
| 💬 LLM / VLM | Transformers · vLLM · PEFT |
| 🐳 Deployment | Docker · NVIDIA Container Toolkit |
| 📈 Experiment Tracking | Weights & Biases · MLflow |
| 🔧 Version Control | Git · GitHub · Git LFS |
| 📦 Data | DVC · Hugging Face Datasets |
| 🧪 Statistics | SciPy · statsmodels · R |
如果本项目对你的学习或研究有所帮助，欢迎引用：
@misc{medicalaiguide2026,
title = {Medical-AI-Guide},
author = {Sun, Yifei and Medical-AI-Guide Contributors},
year = {2026},
howpublished = {\url{https://github.com/diaoquesang/Medical-AI-Guide/}},
}Medical-AI-Guide 是一个持续维护的开源项目。
欢迎：
- 补充优质学习资源
- 推荐值得阅读的论文
- 修复错误或过时内容
- 改进教程
- 分享医学 AI 实践经验
欢迎通过 Issue / Pull Request 参与项目。
📧 Contributor Application: diaoquesang@gmail.com
Build better models. Ask better questions. Do better science.