Back to research
CSE428: Image Processing · 2025

Breast Ultrasound: Multi-Task U-Net

One network that both outlines a lesion in a breast ultrasound image and classifies it as benign, malignant, or normal, trained on the public BUSI dataset.

View on GitHub
[ RES ]

Key results

Dice
0.854
Lesion segmentation, held-out test split
Mean IoU
0.782
Held-out test split
Classification F1
0.863
Three classes, held-out test split
Accuracy
87.5%
Three classes, held-out test split
[ ARC ]

Architecture

A shared encoder feeds two heads: a U-Net decoder with skip connections for the mask, and a pooled classifier on the bottleneck. A weighted sum of the two losses trains both at once.

  1. 01

    Data cleaning

    • Perceptual-hash de-duplication
    • Drop axilla / biopsy images
    • 256 × 256, 60/20/20 split

  2. 02

    Shared encoder

    • Conv blocks + max pooling
    • Bottleneck features

  3. 03

    Two heads

    • U-Net decoder → lesion mask
    • GAP + dense → 3 classes

  4. 04

    Joint loss

    • Segmentation weight 1.0
    • Classification weight 0.5
[ KEY ]

Technical highlights

  1. 01

    Cleaned BUSI before training: removed near-duplicate images with perceptual hashing and dropped axilla and biopsy images that would leak shortcuts.

  2. 02

    Trained segmentation and classification jointly so both tasks share one learned representation.

[ LIM ]

Limits & lessons

  • Metrics come from a single held-out split of a few hundred images; they are coursework results, not clinical validation.
[ STK ]

Stack

  • Python
  • PyTorch
  • Albumentations
  • OpenCV
  • imagehash
  • scikit-learn