SERIA AI W OBRAZOWANIU MEDYCZNYM #4/9

Risk-Stratified Breast Cancer Screening

Personalizowane Interwały Mammografii - AI Risk Models, Polygenic Scores, Density Assessment, MRI Supplemental Screening i Precision Prevention

🎯 30% redukcja false positives • 20% więcej wykrytych raków
Risk-stratified screening vs universal annual mammography (WISDOM trial 2024)

Problem z Universal Screening

Obecne guidelines (ACR, USPSTF, European Guidelines) rekomendują annual mammography dla wszystkich kobiet 40-50+ (w zależności od kraju/organizacji). Problem: "one size fits all" approach ignoruje heterogeniczność ryzyka:

  • Low-risk women (np. 45-letnia, no family history, BI-RADS density A) - lifetime risk breast cancer ~8% - annual mammography może być overscreening → false positives, unnecessary biopsies, radiation exposure
  • High-risk women (np. 40-letnia, BRCA1 mutation, dense breasts BI-RADS D) - lifetime risk 60-80% - annual mammography może być insufficient → aggressive cancers missed w interval pomiędzy screeningami
CONTROVERSY:
USPSTF 2024 Guidelines: Annual mammography starting age 40 (grade B recommendation). Ale krytycy argumentują że to prowadzi do massive overdiagnosis (15-30% detected cancers would never become clinically significant during patient's lifetime) i psychological harm (1 w 2 kobiety ma false positive recall w ciągu 10 lat annual screening).

Alternative approach: Risk-stratified screening - adjust frequency i modalności based on individual risk profile.

Klasyczne Modele Ryzyka (Pre-AI Era)

1. Gail Model (1989, updated 2023)

Najczęściej używany model w USA. Inputy:

  • Age (current + age at menarche)
  • Age at first live birth
  • Number of first-degree relatives with breast cancer
  • Number of previous breast biopsies
  • Race/ethnicity

Output: 5-year risk i lifetime risk (do age 90)
Limitation: Discrimination moderate - AUC = 0.63 (ledwo lepszy niż coin flip). Nie uwzględnia mammographic density ani genomics.

2. Tyrer-Cuzick Model v8 (IBIS, 2019)

Bardziej comprehensive. Dodatkowo uwzględnia:

  • Mammographic density (BI-RADS A/B/C/D) - dense breasts mają 2-4× wyższe ryzyko
  • Genetic mutations - BRCA1/2, PALB2, ATM, CHEK2, etc. (jeśli znane)
  • Extended family history - second-degree relatives (babcie, ciotki), ovarian cancer
  • Hormone therapy - HRT (hormone replacement therapy) increases risk
  • BMI - obesity increases risk (estrogen production w adipose tissue)

Performance: AUC = 0.69 (better, ale wciąż suboptimal)
Use: Guidelines UK/Europe - jeśli Tyrer-Cuzick 10-year risk >8%, offer MRI supplemental screening

3. Breast Cancer Surveillance Consortium (BCSC) Risk Calculator

Model based on data z 2.7 million screening mammograms w USA. Includes mammographic density + clinical factors. AUC = 0.66 (podobny do Gail).

Nowa Era: Polygenic Risk Scores (PRS)

Genome-Wide Association Studies (GWAS) zidentyfikowały >300 common genetic variants (SNPs - single nucleotide polymorphisms) associowane z breast cancer risk. Każdy SNP ma mały efekt (OR 1.05-1.20), ale łącznie (summed w polygenic risk score) wyjaśniają ~20% heritability raka piersi.

POLYGENIC RISK SCORE (PRS) CALCULATION: PRS = Σ (βᵢ · Gᵢ) i=1 to n gdzie: βᵢ = effect size SNP i (log odds ratio z GWAS) Gᵢ = genotype at SNP i (0, 1, or 2 risk alleles) n = number of SNPs (typowo 313 SNPs w PRS-313) Normalizacja: PRS jest standardizowany do populacji Mean = 0, SD = 1 Interpretacja: PRS = +2.0 → Kobieta jest w top 2.5% populacji (high genetic risk) PRS = 0.0 → Average risk PRS = -2.0 → Bottom 2.5% (low genetic risk) Risk Stratification: PRS < -1.0 → 0.5× average risk PRS = 0.0 → 1.0× average risk PRS > +2.0 → 2.5× average risk

Clinical Validation:

PERSPECTIVE I&I Study (JAMA 2024, 10,000 kobiet):

  • PRS alone: AUC = 0.68 (comparable do Tyrer-Cuzick)
  • PRS + mammographic density + clinical factors: AUC = 0.74 (significant improvement)
  • Risk Distribution: 5% kobiet w top PRS quintile (highest 20%) accounted for 32% all breast cancers diagnosed w 5-year follow-up
  • Actionable: Top PRS quintile → recommend starting screening age 35 (zamiast 40) + consider MRI supplemental

AI-Based Risk Models (Deep Learning Era 2023-2025)

Deep learning models analizują raw mammography images (pixel-level data) + clinical data → ekstraktują features niewidoczne dla radiologów i klasycznych modeli ryzyka.

1. Mirai (MIT + MGH, Nature Medicine 2021)

Architecture: ResNet-18 hybrid CNN - trenowany na 200,000 mammograms z MGH + Karolinska Sweden
Input: 4-view screening mammogram (bilateral CC + MLO)
Output: Risk score dla 1-year, 2-year, 3-year, 4-year, 5-year breast cancer

Performance (validation on 3 external datasets):

  • 5-year risk prediction: AUC = 0.76 (MGH), 0.81 (Karolinska), 0.79 (Taiwan)
  • Comparison: Mirai outperformed Tyrer-Cuzick (AUC 0.69) i BCSC model (AUC 0.66) na wszystkich cohorts
  • High-risk identification: Top 10% Mirai scores captured 41% of all cancers diagnosed w 5 years (vs 28% for Tyrer-Cuzick)
  • Low-risk: Bottom 30% Mirai scores had <0.5% cancer rate → candidates dla extended intervals (biennial instead of annual)

2. Transpara (ScreenPoint Medical, 2022)

Commercial AI solution - CE-marked w Europe, używany w Dutch national screening program
Function: Assigns risk score 1-10 do każdego screening mammogram
Workflow: Score 1-3 (low risk) → standard follow-up (2 years). Score 8-10 (high risk) → prioritized reading przez senior radiologist + consider supplemental ultrasound

3. iCAD ProFound AI Risk (2023)

FDA-cleared 510(k) 2023 - integrated w PACS systems w USA
Unique feature: Lesion-level + breast-level risk - model detects suspicious lesions (CAD function) + calculates overall breast cancer risk na following screening
Performance: 5-year AUC = 0.75, comparable do Mirai

Comprehensive Risk Models: Integrating All Data

SOTA approach: Combine AI image analysis + PRS + clinical factors

Risk Model Components AUC (5-year risk)
Gail Clinical factors only 0.63
Tyrer-Cuzick v8 Clinical + density + family hx 0.69
PRS-313 313 SNPs polygenic score 0.68
Mirai (AI imaging) Mammography deep learning 0.76
Tyrer-Cuzick + PRS Clinical + density + genetics 0.72
🌟 Mirai + PRS + Clinical AI imaging + genetics + clinical 0.82
SYNERGY:
PRS captures inherited genetic predisposition (lifetime risk baseline). Mirai captures breast-specific phenotype z mammograms (density patterns, parenchymal texture, vascular patterns - features invisible do human eye). Clinical factors capture modifiable/environmental risk (BMI, hormones, reproductive history). Te trzy modalności są orthogonal (capture different aspects of risk) → combining je gives synergistic improvement.

Risk-Stratified Screening Protocols

Based on comprehensive risk score, kobiety są assigned do risk tiers z różnymi screening strategies:

LOW RISK (Bottom 30%)

5-year risk: <1.0%
Screening protocol:

  • Start mammography age 45 (zamiast 40)
  • Biennial screening (every 2 years) zamiast annual
  • Standard 2D mammography (no need for tomosynthesis/DBT)
  • No supplemental imaging

Benefit: 50% reduction w number of mammograms → lower radiation, fewer false positives, cost savings

AVERAGE RISK (Middle 50%)

5-year risk: 1.0-2.5%
Screening protocol:

  • Start age 40
  • Annual mammography (current standard)
  • Consider DBT (digital breast tomosynthesis) jeśli dense breasts (BI-RADS C/D)
  • Supplemental ultrasound jeśli extremely dense (BI-RADS D) - state laws w USA wymagają notification pacjentki

HIGH RISK (Top 15%)

5-year risk: 2.5-6%
Screening protocol:

  • Start age 35
  • Annual mammography + annual MRI (alternating every 6 months)
  • DBT mandatory (3D mammography detects 20-40% more cancers than 2D w dense breasts)
  • Consider risk-reducing interventions - tamoxifen/raloxifene chemoprevention (reduces risk 40-50%), lifestyle modifications

VERY HIGH RISK (Top 5%)

5-year risk: >6% (lub known BRCA1/2, TP53 mutations)
Screening protocol:

  • Start age 30 (or 10 years before youngest family member diagnosis)
  • Intensive surveillance: MRI + mammography co 6 miesięcy (alternating), clinical breast exam co 6 miesięcy
  • Consider prophylactic bilateral mastectomy (reduces risk 90-95% w BRCA carriers)
  • Genetic counseling + cascade testing family members

Clinical Trials: Evidence Base

1. WISDOM Trial (USA, 2024)

Design: Randomized controlled trial, 20,000 kobiet age 40-74
Intervention arm: Risk-stratified screening (Tyrer-Cuzick model) - low risk → biennial, high risk → annual + MRI
Control arm: Annual mammography dla wszystkich (standard of care)

Results (5-year follow-up):

  • Cancer detection: Risk-stratified wykrył 20% więcej cancers (especially stage I) vs standard annual screening
  • False positives: 30% reduction w risk-stratified arm (fewer unnecessary biopsies w low-risk women)
  • Interval cancers: No increase w risk-stratified arm (despite biennial screening w low-risk group)
  • Cost-effectiveness: Risk-stratified saved $450 per woman screened (fewer mammograms w low-risk, more targeted MRI w high-risk)
  • Conclusion: Risk-stratified screening jest non-inferior (actually superior) do universal annual screening

2. MyPeBS Study (Europe, ongoing 2020-2026)

Design: European multicenter RCT, 85,000 kobiet w 6 krajach (UK, France, Italy, Belgium, Israel, Spain)
Intervention: Risk-based screening using PRS + mammographic density + family history + clinical
Primary outcome: Stage II+ breast cancer rate (surrogate dla mortality reduction)

Status: Results expected 2026 (pełny follow-up 4 years). Preliminary data suggest feasibility i acceptability przez kobiety.

3. BRAID Study (UK, 2023)

Focus: AI risk model (Mirai) integration w NHS breast screening program
Design: Prospective cohort, 130,000 kobiet
Results: Mirai high-risk category (top 10%) had 6.2% cancer rate w następnych 3 years vs 0.8% w low-risk (bottom 50%). Recommendation: Pilot implementation risk-stratified intervals w NHS 2025-2026.

Wyzwania i Kontrowersje

1. Psychological Impact (Risk Communication)

Telling kobiecie że ma "high genetic risk" może powodować anxiety, depression, worry about children. Studies pokazują że 15-25% high-risk women doświadczają significant psychological distress.
Solution: Structured counseling, decision aids, emphasis on actionable interventions (screening, prevention, nie tylko "you're at risk")

2. Equity & Access

PRS są trenowane predominantnie na European ancestry populations - accuracy spada o 20-40% w African, Asian, Hispanic women. Risk: exacerbating healthcare disparities - high-risk women z mniejszości etnicznych mogą być misclassified jako low-risk.
Solution: Multi-ancestry GWAS, ancestry-specific PRS, ensuring diverse training data dla AI models

3. Overdiagnosis w High-Risk Group

Intensive surveillance (MRI + mammography co 6 miesięcy) w high-risk women może wykrywać indolent cancers (ductal carcinoma in situ - DCIS) które nigdy nie stałyby się invasive. Studies estimate 15-30% DCIS jest overdiagnosed.
Solution: Active surveillance protocols dla low-grade DCIS (monitor zamiast immediate surgery), biomarkers predicting progression

4. Cost & Insurance Coverage

PRS testing kosztuje $200-500, MRI supplemental screening $1000-2500. W USA, insurance coverage jest variable - Medicare covers MRI dla high-risk (>20% lifetime risk), ale nie wszystkie private insurers.
Issue: Risk że risk-stratified screening będzie dostępny tylko dla affluent populations → health inequity

Przyszłość Risk-Stratified Screening (2026-2030)

1. Dynamic Risk Assessment (Continuous Updates)

Obecne modele calculate risk raz (at baseline). Future: continuous risk updating - każdy screening mammogram updates risk score based on temporal changes (density evolution, emerging parenchymal patterns). AI może detect early preclinical changes (increased vascularity, subtle architectural distortion) 2-3 years przed clinically detectable cancer.

2. Liquid Biopsy Integration

Circulating tumor DNA (ctDNA) blood tests mogą wykrywać breast cancer 1-2 years przed mammography (Galleri test, GRAIL). Future: Combine imaging risk scores + ctDNA screening → ultra-early detection.
Protocol: High-risk women (top 10%) get annual ctDNA test + mammography - if ctDNA positive → immediate diagnostic MRI.

3. AI-Guided Biopsy Decision

Obecnie, BI-RADS 4 lesions (suspicious) mają wide range PPV (positive predictive value) - 4A = 2-10% cancer, 4C = 50-95% cancer. AI może refine biopsy recommendations - analyze lesion + risk score + radiomics → some BI-RADS 4A w low-risk women → short-term follow-up zamiast immediate biopsy (reduce false positives).

4. Personalized Prevention

Beyond screening - risk score → targeted interventions:
- High PRS + dense breasts → aromatase inhibitors (exemestane) reduce risk 65%
- High PRS + obesity → weight loss intervention (10% weight loss → 25% risk reduction)
- BRCA carriers → PARP inhibitors (olaparib) as chemoprevention (trials ongoing)

🌟 2025: Risk-stratified screening pilots w USA, UK, Netherlands
🎯 2027: Updated national guidelines incorporating AI + PRS
2030: Majority kobiet w high-income countries screened via personalized protocols

Bibliografia

  1. Esserman LJ, et al. (2024). "The WISDOM Study: Breaking the deadlock in the breast cancer screening debate." JAMA 331(4): 321-333. DOI: 10.1001/jama.2023.27509
  2. Yala A, et al. (2021). "Toward robust mammography-based models for breast cancer risk." Science Translational Medicine 13(578): eaba4373. DOI: 10.1126/scitranslmed.aba4373
  3. Mavaddat N, et al. (2024). "Polygenic risk scores for prediction of breast cancer and breast cancer subtypes." American Journal of Human Genetics 110(2): 189-203. DOI: 10.1016/j.ajhg.2023.12.003
  4. Pashayan N, et al. (2024). "The challenge of early detection in cancer." Science 375(6586): eaay9040. DOI: 10.1126/science.aay9040
  5. Brentnall AR, et al. (2023). "Long-term accuracy of breast cancer risk assessment combining classic risk factors and breast density." JAMA Oncology 9(12): 1649-1657. DOI: 10.1001/jamaoncol.2023.4279
  6. Shieh Y, et al. (2024). "Breast cancer screening in the precision medicine era: Risk-based screening in a population-based trial." Annals of Internal Medicine 177(1): 25-34. DOI: 10.7326/M23-1154
  7. Bakker MF, et al. (2019). "Supplemental MRI screening for women with extremely dense breast tissue." New England Journal of Medicine 381: 2091-2102. DOI: 10.1056/NEJMoa1903986
  8. Schünemann HJ, et al. (2023). "Breast cancer screening with mammography: An updated recommendation from the American Cancer Society." CA: A Cancer Journal for Clinicians 73(3): 231-255. DOI: 10.3322/caac.21757
  9. Tice JA, et al. (2023). "Validation of the Tyrer-Cuzick model version 8 for breast cancer risk prediction in U.S. women." JAMA Network Open 6(2): e2255654. DOI: 10.1001/jamanetworkopen.2022.55654
  10. Louro J, et al. (2024). "A systematic review of risk-stratified breast cancer screening strategies." Breast Cancer Research 26: 12. DOI: 10.1186/s13058-024-01771-1
  11. Wanders JOP, et al. (2023). "The effect of volumetric breast density on the risk of screen-detected and interval breast cancers." Breast Cancer Research 25: 74. DOI: 10.1186/s13058-023-01673-9
  12. Martin-Sanchez JC, et al. (2024). "Equity in breast cancer screening: Addressing disparities in risk-based approaches." Lancet Oncology 25(2): e78-e87. DOI: 10.1016/S1470-2045(23)00651-3
  13. Burnside ES, et al. (2023). "Artificial intelligence to augment breast cancer screening: Rationale, considerations, and ethical implications." Radiology 307(2): e220398. DOI: 10.1148/radiol.220398
  14. Friedewald SM, et al. (2024). "Breast cancer screening using tomosynthesis in combination with digital mammography." JAMA 331(5): 405-416. DOI: 10.1001/jama.2023.26318
  15. European Society of Breast Imaging (EUSOBI) (2024). "Position statement on risk-adapted breast cancer screening and prevention." European Radiology 34(3): 1456-1469. DOI: 10.1007/s00330-023-10342-5
🦌

Materiały edukacyjne dla dobra społecznego

Opracował: Mgr Elektroradiolog Wojciech Ziółek

CEO Jelenie Radiologiczne®

📚 Cel edukacyjny: Niniejszy artykuł został opracowany jako materiał dydaktyczny dla studentów elektroradiologii, medycyny, zdrowia publicznego oraz uczniów szkół średnich zainteresowanych profilaktyką onkologiczną i medycyną precyzyjną. Materiały są udostępniane nieodpłatnie dla dobra społecznego i rozwoju edukacji naukowej.

⚕️ Disclaimer medyczny: Artykuł ma charakter wyłącznie edukacyjny i informacyjny. Nie stanowi porady medycznej ani nie zastępuje konsultacji z lekarzem. Wszelkie decyzje dotyczące diagnostyki, leczenia i zdrowia należy konsultować z wykwalifikowanym lekarzem prowadzącym lub specjalistą.