U-Net Segmentation for Dermatology Images

Deep learning for medical image segmentation

Overview

A computer vision project implementing U-Net architecture for automated lesion segmentation in dermatological images, with focus on handling small-sample datasets common in medical imaging.

Clinical Significance

Dermatological image segmentation enables:

  • Automated lesion boundary detection
  • Size and morphology quantification
  • Follow-up comparison and monitoring
  • Melanoma risk assessment
  • Objective treatment response evaluation

Key Achievements

  • Small-Sample Optimization: High performance with <500 labeled images
  • Class-Specific IoU Breakdown: Separate performance metrics for skin, lesion, background
  • Data Augmentation: Custom strategies for medical image domain
  • Generalization: Cross-validation on diverse skin types and lesion morphologies

Methodology

Dataset Preparation

  • Source: Public dermatology image datasets (ISIC Archive)
  • Preprocessing:
    • Standardization to 512×512 resolution
    • Histogram equalization for consistency
    • Manual annotation verification
  • Size: 450 training, 100 validation, 100 test images

Data Augmentation Strategy

  • Geometric Transformations: Rotation (0-90°), flipping, elastic deformation
  • Intensity Modifications: Brightness/contrast adjustment, Gaussian blur
  • Domain Shifts: Color space variations, synthetic shadow/lighting changes
  • Mixup: Blending augmented samples for improved robustness

Model Architecture

Input: 512×512 RGB Image
  ↓
Encoder: 4-level downsampling (64 → 512 filters)
  ↓
Bottleneck: Convolutional blocks
  ↓
Decoder: 4-level upsampling with skip connections
  ↓
Output: 512×512 binary segmentation mask

Training Configuration

  • Optimizer: Adam (learning rate 0.001)
  • Loss: Dice Loss + Cross-Entropy (weighted combination)
  • Batch Size: 16
  • Epochs: 200 with early stopping
  • Validation: 5-fold cross-validation

Results

Segmentation Performance

  • Mean IoU: 0.87 ± 0.05
  • Dice Coefficient: 0.92 ± 0.03
  • Sensitivity: 0.89 (lesion detection)
  • Specificity: 0.95 (false positive suppression)

Class-Specific Metrics

  • Skin (Background): IoU = 0.94
  • Lesion: IoU = 0.87
  • Boundary Accuracy: 2-3 pixels median error

Clinical Applications

  • Melanoma screening support systems
  • Treatment response monitoring (chemotherapy, immunotherapy)
  • Surgical planning and margin definition
  • Population-level skin disease epidemiology

Future Work

  • Multi-class segmentation (lesion type classification)
  • 3D reconstruction from multiple views
  • Integration with dermoscopy features
  • Clinical trial deployment

Timeline

  • Start: February 2021
  • End: June 2021
  • Duration: 5 months