Medical Imaging GuideSAFE

🔬 A curated collection of 23,000+ agent skills for empirical research across 8 social science disciplines. | 精选 23,000+ AI Agent 技能库,覆盖8大社会科学学科的实证研究。CoPaper.AI 20分钟完成一篇可复现的规范实证论文,并支持用户上传 Skills。-- Maintained by CoPaper.AI from Stanford REAP.

The instruction file, verbatim from the audited commit — this is the text the model reads, and the surface the audit's instruction layer examines. Quoted here so you can judge it without cloning anything.

---
name: medical-imaging-guide
description: "Medical image analysis with deep learning for research applications"
metadata:
  openclaw:
    emoji: "🔬"
    category: "domains"
    subcategory: "biomedical"
    keywords: ["medical imaging", "deep learning", "image segmentation", "radiology AI", "pathology", "convolutional neural networks"]
    source: "wentor-research-plugins"
---

# Medical Imaging Guide

A skill for applying deep learning to medical image analysis in research settings. Covers common imaging modalities, preprocessing pipelines, architecture selection for classification and segmentation tasks, handling small datasets with transfer learning and data augmentation, evaluation metrics specific to medical imaging, and regulatory and ethical considerations for clinical translation.

## Imaging Modalities and Data Characteristics

### Common Modalities in Research

```
Modality Overview:

X-ray / Radiography:
  - 2D grayscale images
  - Resolution: typically 2000x2000 to 4000x4000 pixels
  - Format: DICOM (.dcm)
  - Common tasks: pneumonia detection, fracture detection,
    cardiomegaly screening
  - Dataset examples: CheXpert, MIMIC-CXR, NIH ChestX-ray14

CT (Computed Tomography):
  - 3D volumetric data (stack of 2D slices)
  - Resolution: 512x512 per slice, 50-500+ slices
  - Format: DICOM series, NIfTI (.nii.gz)
  - Common tasks: lung nodule detection, organ segmentation,
    COVID-19 screening
  - Dataset examples: LUNA16, DeepLesion, TotalSegmentator

MRI (Magnetic Resonance Imaging):
  - 3D volumetric, multiple sequences (T1, T2, FLAIR, DWI)
  - Resolution: 256x256 to 512x512 per slice
  - Format: DICOM, NIfTI
  - Common tasks: brain tumor segmentation, cardiac analysis,
    knee injury classification
  - Dataset examples: BraTS, ACDC, fastMRI

Histopathology:
  - Whole slide images (WSI), extremely large
  - Resolution: 100,000x100,000+ pixels at highest magnification
  - Format: SVS, TIFF, NDPI (vendor-specific)
  - Common tasks: cancer grading, mitosis detection,
    tissue classification
  - Dataset examples: Camelyon16/17, TCGA, PANDA

Retinal Imaging (Fundoscopy / OCT):
  - 2D color fundus or 3D OCT volumes
  - Common tasks: diabetic retinopathy grading, glaucoma detection
  - Dataset examples: EyePACS, MESSIDOR, REFUGE
```

## Preprocessing Pipeline

### Standard Preprocessing Steps

```python
import numpy as np

def preprocess_medical_image(image, modality="xray"):
    """
    Standard preprocessing pipeline for medical images.

    Steps vary by modality but typically include:
    1. Intensity normalization
    2. Resizing/resampling
    3. Windowing (for CT)
    4. Artifact removal
    """
    if modality == "ct":
        # CT windowing: map Hounsfield Units to display range
        # Lung window: center=-600, width=1500
        # Soft tissue: center=40, width=400
        window_center = -600
        window_width = 1500
        lower = window_center - window_width // 2
        upper = window_center + window_width // 2
        image = np.clip(image, lower, upper)
        image = (image - lower) / (upper - lower)

    elif modality == "xray":
        # Normalize to [0, 1] range
        image = image.astype(np.float32)
        image = (image - image.min()) / (image.max() - image.min() + 1e-8)

    elif modality == "mri":
        # Z-score normalization (per-volume)
        # Exclude background (zeros) from statistics
        mask = image > 0
        if mask.any():
            mean_val = image[mask].mean()
            std_val = image[mask].std()
            image = (image - mean_val) / (std_val + 1e-8)

    return image


def resize_with_spacing(image, original_spacing, target_spacing):
    """
    Resample 3D medical image to uniform voxel spacing.
    Essential for CT/MRI where slice thickness varies.

    Args:
        image: 3D numpy array
        original_spacing: (z, y, x) voxel sizes in mm
        target_spacing: desired (z, y, x) voxel sizes in mm
    """
    from scipy.ndimage import zoom

    resize_factor = [
        orig / target
        for orig, target in zip(original_spacing, target_spacing)
    ]
    resampled = zoom(image, resize_factor, order=1)
    return resampled
```

## Model Architecture Selection

### Task-Specific Architectures

```
Architecture recommendations by task:

IMAGE CLASSIFICATION (diagnosis, grading):
  - ResNet-50/101: reliable baseline, well-understood
  - EfficientNet-B4/B5: better accuracy-efficiency tradeoff
  - Vision Transformer (ViT): strong with large datasets
  - DenseNet-121: popular for chest X-ray (CheXNet heritage)

  Transfer learning approach:
    1. Start with ImageNet pretrained weights
    2. Replace final classifier layer
    3. Fine-tune with low learning rate (1e-4 to 1e-5)
    4. Use gradual unfreezing (train head, then all layers)

IMAGE SEGMENTATION (organ/lesion delineation):
  - U-Net: gold standard for medical segmentation
  - nnU-Net: self-configuring U-Net (state-of-the-art framework)
  - Attention U-Net: U-Net with attention gates
  - TransUNet: hybrid CNN-Transformer architecture
  - MONAI: framework with pre-built medical imaging models

  For 3D volumes:
    - 3D U-Net: full volumetric processing
    - V-Net: 3D with dice loss
    - 2.5D approach: adjacent slices as multi-channel input

OBJECT DETECTION (lesion localization):
  - YOLO variants: fast inference, suitable for screening
  - Faster R-CNN: higher accuracy, slower
  - RetinaNet: handles class imbalance (focal loss)
  - DETR: transformer-based, no anchor boxes needed
```

## Handling Small Datasets

### Data Augmentation for Medical Images

```python
def get_medical_augmentation_pipeline():
    """
    Medical image augmentation strategy.

    Key differences from natural image augmentation:
    - Preserve anatomical plausibility
    - Avoid color jitter (intensity has diagnostic meaning)
    - Use elastic deformations (mimic anatomical variability)
    - Apply carefully (aggressive augmentation can hurt)
    """
    import albumentations

SAFEgrade B · trust 89/100 Nothing in the source contradicts what it says it does. Grade A is reserved for packages that have also passed the behavioural sandbox.

What the source does

Filesystem

none-observed

Network

none-observed

Shell

none-observed

Dependencies

pinned

Secrets in source

none-found

Findings (0)

No findings outside the package's declared scope.

Gates applied: no_behavioural_pass.