← Lectures·L16

Deep Learning in Breast MRI

Kapsner · 1 concepts · 6 questions

Concept 1 / 1

Key Concepts to Memorize

Breast cancer screening context:


  • Most common cancer in women
  • X-ray mammography: screening every 2 years for women aged 50–75
  • Mammography screening → -30% decrease in breast cancer mortality (Swedish Two-County Trial)

3 main challenges in breast cancer screening:


1. False-positive findings → unnecessary biopsies

2. False-negative findings → missed cancers ("interval cancer")

3. Over-diagnosis → over-treatment


Why add MRI to screening?


  • MRI has much higher cancer detection rate (CDR/1000: 16.5 with MRI vs. 5–6 with mammography alone)
  • But MRI limitations: high cost, long duration, requires contrast agents, limited availability

EUSOBI Recommendations 2024:


  • Regular mammography remains the mainstay of breast cancer screening
  • High-risk women + extremely dense breast tissue → use MRI
  • Women should actively participate in personalized screening decisions

4 DL applications in breast MRI (Kapsner's 4 topics):


1. Quality Assurance — automated artifact detection in MRI-derived MIPs

2. Virtual Contrast Enhancement (vCE) — generate contrast-enhanced appearance from contrast-free MRI

3. Lesion Detection — AI-powered CAD system for breast DWI MRI

4. Additional/Incidental Findings — detect pathologies outside region of interest (e.g., aortic aneurysm)


Maximum Intensity Projection (MIP):


  • Step 1: subtract pre-contrast from post-contrast 3D image → only contrast-enriched areas remain
  • Step 2: project maximum intensities along one axis → 2D overview image
  • 66.9% of MIPs have artifacts → DL achieves ~86-94% AUROC in artifact detection

Virtual Contrast Enhancement (vCE):


  • Goal: predict contrast enhancement from contrast-free MRI sequences (T1, T2, DWI)
  • Avoids risks of contrast agents: allergic reactions, kidney failure, environmental contamination
  • GAN produces more realistic-looking vCE images than U-Net (confirmed by Turing test)

Thoracic Aortic Aneurysm (incidental finding):


  • "Silent killer": >95% asymptomatic; rupture → >90% mortality
  • AI detected aneurysms at ~3.5× higher rate than routine clinical reporting
  • Women have +40% probability to die from thoracic aortic aneurysm vs. men