Skip to content
Guides3 min read

使用 Potato 進行醫學影像標註

使用 Potato 標準影像標註功能進行醫學影像標註的最佳實踐。

Potato Team

Potato 的影像標註功能可用於標註醫學影像,如 X 光片、皮膚病變照片或其他標準格式(PNG、JPG)的臨床影像。本指南涵蓋了設定醫學影像標註項目的最佳實踐。

重要注意事項

在啟動任何醫學標註項目之前:

  1. 去標識化:確保在標註前已從影像中移除所有受保護的健康資訊(PHI)
  2. 資料處理:使用機構的安全基礎設施儲存和處理醫學資料
  3. 訪問控制:通過現有的身份驗證系統管理標註者訪問
  4. 合規性:與 IRB 和合規團隊合作,滿足機構要求

注意:Potato 是一個標註工具,不包含內建的 HIPAA 合規功能、DICOM 檢視器或專用醫學影像功能。這些需求應在基礎設施層面解決。

基本醫學影像標註設定

以下是為標準醫學影像配置 Potato 的方法:

yaml
annotation_task_name: "Medical Image Annotation"
 
data_files:
  - path: data/medical_images.json
    type: json
 
annotation_schemes:
  - annotation_type: image_annotation
    tools: [bbox]
    name: lesions
    labels:
      - name: primary
        color: "#EF4444"
        description: "Primary finding"
 
      - name: secondary
        color: "#F97316"
        description: "Secondary finding"
 
      - name: artifact
        color: "#6B7280"
        description: "Artifact or noise"
 
  - annotation_type: radio
    name: image_quality
    description: "Overall image quality?"
    labels:
      - name: diagnostic
        label: "Diagnostic Quality"
      - name: limited
        label: "Limited Quality"
      - name: non_diagnostic
        label: "Non-Diagnostic"

X 光片和臨床照片標註

對於已匯出為標準影像格式的 X 光片或臨床照片:

yaml
annotation_task_name: "X-Ray Findings Annotation"
 
data_files:
  - path: data/xray_images.json
    type: json
 
annotation_schemes:
  - annotation_type: image_annotation
    tools: [polygon]
    name: abnormalities
    labels:
      - name: opacity
        color: "#EF4444"
        description: "Pulmonary opacity"
 
      - name: consolidation
        color: "#F97316"
        description: "Consolidation"
 
      - name: nodule
        color: "#EAB308"
        description: "Pulmonary nodule"
 
      - name: effusion
        color: "#3B82F6"
        description: "Pleural effusion"
 
  - annotation_type: multiselect
    name: findings
    description: "Select all findings present"
    labels:
      - name: normal
        label: "No acute findings"
      - name: pneumonia
        label: "Pneumonia"
      - name: atelectasis
        label: "Atelectasis"
      - name: pneumothorax
        label: "Pneumothorax"
      - name: fracture
        label: "Fracture"

皮膚病變標註

用於皮膚科影像:

yaml
annotation_task_name: "Dermoscopy Annotation"
 
data_files:
  - path: data/skin_lesions.json
    type: json
 
annotation_schemes:
  - annotation_type: image_annotation
    tools: [polygon]
    name: lesion_boundary
    labels:
      - name: lesion
        color: "#EF4444"
        description: "Lesion boundary"
 
  - annotation_type: multiselect
    name: dermoscopic_features
    description: "Select all features present"
    labels:
      - name: pigment_network
        label: "Pigment Network"
      - name: dots_globules
        label: "Dots/Globules"
      - name: streaks
        label: "Streaks"
      - name: blue_white_veil
        label: "Blue-White Veil"
      - name: regression
        label: "Regression Structures"
      - name: vascular
        label: "Vascular Structures"
 
  - annotation_type: radio
    name: diagnosis
    description: "Most likely diagnosis?"
    labels:
      - name: benign_nevus
        label: "Benign Nevus"
      - name: seborrheic_keratosis
        label: "Seborrheic Keratosis"
      - name: basal_cell
        label: "Basal Cell Carcinoma"
      - name: melanoma
        label: "Melanoma"
      - name: other
        label: "Other"
 
  - annotation_type: likert
    name: confidence
    description: "Diagnostic confidence"
    size: 5
    min_label: "Low"
    max_label: "High"

視網膜影像標註

用於眼底照片和其他視網膜影像:

yaml
annotation_task_name: "Fundus Image Annotation"
 
data_files:
  - path: data/fundus_images.json
    type: json
 
annotation_schemes:
  - annotation_type: image_annotation
    tools: [polygon]
    name: anatomical_structures
    labels:
      - name: optic_disc
        color: "#FDE68A"
        description: "Optic disc boundary"
 
      - name: fovea
        color: "#A78BFA"
        description: "Fovea region"
 
      - name: macula
        color: "#93C5FD"
        description: "Macular region"
 
  - annotation_type: image_annotation
    tools: [polygon]
    name: pathology
    labels:
      - name: hemorrhage
        color: "#EF4444"
        description: "Retinal hemorrhage"
 
      - name: exudate
        color: "#FCD34D"
        description: "Hard/soft exudate"
 
      - name: microaneurysm
        color: "#F97316"
        description: "Microaneurysm"
 
  - annotation_type: radio
    name: dr_grade
    description: "Diabetic retinopathy grade"
    labels:
      - name: none
        label: "No DR"
      - name: mild
        label: "Mild NPDR"
      - name: moderate
        label: "Moderate NPDR"
      - name: severe
        label: "Severe NPDR"
      - name: proliferative
        label: "PDR"

資料格式

以 JSON 格式準備影像資料:

json
[
  {
    "id": "case_001",
    "image": "images/case_001.png",
    "metadata": {
      "body_part": "chest",
      "modality": "xray"
    }
  },
  {
    "id": "case_002",
    "image": "images/case_002.png",
    "metadata": {
      "body_part": "chest",
      "modality": "xray"
    }
  }
]

注意:影像必須為標準的 Web 相容格式(PNG、JPG 等)。DICOM 檔案需要在使用 Potato 之前轉換為標準影像格式。

最佳實踐

  1. 使用合格的標註者:醫學標註需要臨床專業知識——確保標註者接受過適當培訓
  2. 預處理影像:將專用格式(DICOM 等)轉換為標準影像格式,並確保 PHI 已移除
  3. 多標註者:讓多個標註者標註每張影像以衡量標註者間一致性
  4. 清晰的指南:提供針對臨床領域的詳細標註指南
  5. 品質檢查:包含金標準案例以監控標註者準確性
  6. 機構合規:與合規團隊合作確保資料處理符合要求

侷限性

Potato 提供通用影像標註,不包含:

  • 原生 DICOM 檔案支援或 DICOM 檢視器
  • 全切片影像(WSI)檢視器及多解析度縮放
  • 內建 HIPAA 合規或審計日誌
  • 醫學資質驗證
  • 專業放射學窗位工具

如需這些功能,請考慮預處理資料或將 Potato 與專業醫學影像基礎設施整合。


有關影像標註的更多資訊,請參閱 影像標註