feature-extraction

---

name: feature-extraction description: Extracts vector embeddings for fairness analysis using MobileNetV3 with FairDisCo disentanglement license: MIT compatibility: opencode metadata: audience: developers workflow: clinical-pipeline

What I do

I extract feature embeddings from segmented skin regions using MobileNetV3 with FairDisCo (Fair Disentangled Continual) disentanglement. This ensures features are suitable for downstream fairness analysis across demographic groups.

When to use me

Use this when:

• Segmentation is complete and you need feature vectors
• You need embeddings for lesion classification
• Fairness analysis requires disentangled feature representations

Key Concepts

• MobileNetV3: Efficient CNN for feature extraction
• FairDisCo: Disentanglement method for fair representations
• Feature Vector: Embedding suitable for classification and similarity
• features_extracted: State flag after extraction complete

Source Files

• services/vision.ts: Feature extraction implementation
• types.ts: FeatureMetadata interface

Code Patterns

• Run MobileNetV2/V3 on segmented skin regions
• Apply FairDisCo disentanglement for fair representations
• Return feature vector and bias metrics

Operational Constraints

• Must use tf.tidy() or explicit dispose() for all tensors
• FairDisCo ensures demographic bias is minimized
• Feature quality directly impacts lesion detection accuracy