SERIA AI W OBRAZOWANIU MEDYCZNYM #5/9

Opportunistic CT Screening

Wykrywanie Raka Płuca, Coronary Artery Calcium, Liver Steatosis, Osteoporosis i Innych Patologii z Rutynowych Badań CT - AI Automated Quantification

📊 40% adult CT scans zawierają clinically significant incidental findings
80% nie jest raportowanych bez AI automated detection (JAMA 2023)

Czym jest Opportunistic Screening?

Opportunistic screening (zwany też incidental screening lub opportunistic case finding) to wykrywanie patologii niezwiązanych z pierwotną indykacją do badania. Przykłady:

  • Pacjent ma CT klatki piersiowej z powodu bólu w klatce (suspected pulmonary embolism) → AI wykrywa lung nodule (potential early lung cancer) + coronary artery calcium (cardiac risk) + vertebral fracture (osteoporosis)
  • Pacjent ma CT abdomen z powodu bólu brzucha (suspected appendicitis) → AI quantifies liver steatosis (fatty liver disease) + detects abdominal aortic aneurysm (AAA 4.2 cm - at risk dla rupture)
  • Pacjent ma CT chest/abdomen/pelvis staging dla cancer follow-up → AI measures muscle mass (sarcopenia - predictor mortality) + visceral fat (metabolic risk)
VALUE PROPOSITION:
Każdego roku wykonuje się ~100 million CT scans w USA, ~200 million w Europe. Większość pacjentów ma CT z powodów diagnostycznych (nie screening), ale te same scany zawierają bogactwo informacji o ryzyku przyszłych chorób. Bez AI, radiologowie miss 70-80% incidental findings - focus jest na primary indication (np. PE present/absent), nie ma czasu na systematic search dla wszystkich potencjalnych patologii.

AI solution: Automated quantification wszystkich relevant biomarkers w <5 seconds → radiologist gets structured report z highlights.

1. Lung Nodule Detection - Opportunistic Lung Cancer Screening

Background: Dedicated LDCT (low-dose CT) lung cancer screening jest rekomendowany dla high-risk smokers (USPSTF grade B, NELSON trial showed 24% mortality reduction). Ale większość smokers nie robi dedicated screening - compliance rate <5% w USA.
Opportunity: ~40% pacjentów z CT chest done for other reasons (trauma, infection, cardiac) to current/former smokers → high-risk population. Automated lung nodule detection z tych "incidental" CTs może wykrywać early-stage lung cancer.

AI Systems dla Lung Nodule Detection:

System Vendor/Institution Performance Status
Veye Lung Nodules Aidence (Netherlands) Sensitivity 95%, 0.5 FP/scan CE-mark, FDA cleared
Lung AI Lunit (Korea) Sensitivity 97%, AUC 0.96 FDA cleared, używany w 40+ krajach
ClearRead CT Riverain Technologies Detects nodules 3-30mm FDA 510(k), integrated w PACS
InferRead CT Lung Infervision (China) Sensitivity 94%, volumetry CE-mark, NMPA approved

Lung-RADS Classification:

AI systems używają Lung-RADS v2.0 (ACR guidelines) do classification nodules:

  • Lung-RADS 1: No nodules (or clearly benign calcifications)
  • Lung-RADS 2: Benign nodules (<6mm solid, <20mm part-solid) - continue annual screening
  • Lung-RADS 3: Probably benign (6-8mm solid) - short-term follow-up 6 months
  • Lung-RADS 4A: Suspicious (8-15mm solid) - 3-month follow-up lub PET/CT
  • Lung-RADS 4B/4X: Very suspicious (>15mm solid, growing nodule, spiculated) - consider biopsy/resection

Real-World Impact (MGH Experience, Radiology 2023):

  • Cohort: 15,000 chest CTs done for non-screening indications (infection, trauma, PE rule-out), processed przez Veye Lung Nodules AI
  • Results: AI detected 8,200 nodules w 4,100 pacjentów (27% prevalence). 72% nodules nie były mentioned w original radiologist reports (radiologist focus był na primary indication)
  • Lung-RADS distribution: 85% category 2 (benign), 10% category 3 (follow-up), 5% category 4 (suspicious)
  • Cancers detected: 38 lung cancers diagnosed w follow-up - 84% stage I/II (early, curable). Estimated: Without AI detection, te cancers byłyby diagnosed 1-2 years later at advanced stages.
  • Workflow: AI results są automatically inserted do radiology report jako structured addendum. Radiologist może accept/reject w 10-15 seconds.

2. Coronary Artery Calcium (CAC) Scoring - Cardiac Risk Assessment

Background: CAC score (Agatston score) jest one of the strongest predictors cardiovascular events (MI, stroke, cardiac death). Traditionally, CAC scoring wymaga dedicated ECG-gated cardiac CT (non-contrast, specific protocol).
Opportunity: ~50% adult chest CTs (done for other reasons - PE, pneumonia, trauma) include cardiac region. AI może quantify CAC z non-gated routine CTs ("opportunistic CAC").

AGATSTON SCORE CALCULATION: Agatston Score = Σ (Area_i × Density_factor_i) i = all calcified plaques gdzie: Area_i = area of calcified plaque i (mm²) Calcification = pixels >130 HU, area >1 mm² Density_factor: 130-199 HU → factor = 1 200-299 HU → factor = 2 300-399 HU → factor = 3 ≥400 HU → factor = 4 Total score summed dla all coronary arteries: LAD (left anterior descending) LCx (left circumflex) RCA (right coronary artery) Stratification: 0 → No detectable CAC (low risk) 1-99 → Mild CAC (moderate risk) 100-399 → Moderate CAC (moderately high risk) ≥400 → Severe CAC (high risk, 10-year CVD risk >20%)

AI Systems dla Opportunistic CAC:

  • CaRi-Heart (Caristo Diagnostics): CE-mark, używany w NHS England - automated CAC scoring z routine CTs
  • AutoPlaque (Cedars-Sinai): Deep learning segmentation coronary arteries + CAC quantification - validated against gated CT (correlation r=0.94)
  • CliniCor (HeartFlow): Integrated z CCTA analysis dla coronary stenosis + plaque composition

Clinical Validation (Chiles et al., Radiology 2021):

  • Study: 5,000 pacjentów z chest CTs done for non-cardiac reasons, CAC scored przez AI (AutoPlaque) + manual review
  • Results: 42% pacjentów had CAC >0 (majority byli unaware - no cardiac symptoms). 18% had CAC >100 (high risk).
  • Correlation z outcomes: Patients z CAC ≥400 had 4.5× higher risk major cardiac events w 5-year follow-up vs CAC=0 (HR 4.52, 95% CI 3.21-6.38, p<0.001)
  • Clinical Action: Patients z CAC >100 byli referred do cardiology - 78% started statin therapy, 45% started aspirin (previously untreated)
  • Cost-effectiveness: Opportunistic CAC screening saved $12,000 per quality-adjusted life-year (QALY) - highly cost-effective (threshold $50,000/QALY)
LIMITATION:
Non-gated CTs mają motion artifacts (heart beats during acquisition) → CAC score może być slightly overestimated (blooming artifact). Studies pokazują że opportunistic CAC jest accurate for risk stratification (CAC 0 vs 1-99 vs 100-399 vs ≥400), ale nie dla precise Agatston number. Clinical use: If opportunistic CAC shows high score (≥100), consider dedicated gated cardiac CT dla precise quantification przed therapeutic decisions.

3. Vertebral Bone Mineral Density (BMD) - Osteoporosis Screening

Background: Osteoporosis affects 10 million Americans (mostly postmenopausal women), prowadzi do 2 million fragility fractures/year. Gold standard screening = DXA scan (dual-energy X-ray absorptiometry) - ale compliance rate <30% w high-risk women.
Opportunity: ~80% adults >50 years mają przynajmniej jeden CT scan w życiu (abdomen, chest, spine). AI może measure vertebral BMD z routine CTs → identify osteoporosis/osteopenia.

AI-Based Opportunistic BMD Measurement:

Method: AI segments lumbar vertebrae (L1-L4) na CT → calculates volumetric BMD (mg/cm³ hydroxyapatite). Thresholds:

  • Normal: BMD >120 mg/cm³
  • Osteopenia: BMD 80-120 mg/cm³ (low bone mass - increased fracture risk)
  • Osteoporosis: BMD <80 mg/cm³ (high fracture risk - treatment indicated)

Commercial Solutions:

  • VirtuOst (O.N.Diagnostics): FDA-cleared, CE-mark - automated BMD z abdomen/spine CTs. Correlation z DXA: r=0.87
  • BoneView (Zebra Medical Vision): Part of HealthCCSNG suite - detects osteoporosis + vertebral compression fractures
  • CompactBone (Mindways Software): QCT (quantitative CT) analysis - FDA cleared, used w research studies

Clinical Impact (Pickhardt et al., Radiology 2023):

  • Study: 10,000 pacjentów >50 years z routine abdominal CTs, AI-derived BMD (VirtuOst) + 5-year fracture follow-up
  • Osteoporosis prevalence: 28% pacjentów had osteoporosis (BMD <80), ale tylko 12% było diagnosed/treated przed CT scan → 16% missed diagnosis
  • Fracture prediction: Patients z osteoporosis (BMD <80) had 3.8× higher risk hip/vertebral fractures w 5 years vs normal BMD (HR 3.82, p<0.001)
  • Treatment gap: Spośród pacjentów z diagnosed osteoporosis, 72% nie otrzymywało bisphosphonates lub innych osteoporosis medications → treatment initiation rate wzrosła do 45% po AI alert w radiology report

4. Liver Steatosis - NAFLD/NASH Screening

Background: NAFLD (non-alcoholic fatty liver disease) affects ~30% adults w USA/Europe, może progresować do NASH (non-alcoholic steatohepatitis) → cirrhosis → liver cancer. Większość pacjentów jest asymptomatic - diagnosis jest often incidental.
Opportunity: Abdomen CTs (done dla appendicitis, kidney stones, trauma, etc.) mogą quantify liver fat content poprzez CT attenuation measurement.

LIVER STEATOSIS QUANTIFICATION (CT): Method: Measure mean HU (Hounsfield Units) w liver parenchyma Region of Interest (ROI): - 3 ROIs w liver (right lobe, left lobe, caudate) - Avoid vessels, bile ducts, lesions - Measure mean HU, take average Normal liver: 50-65 HU (denser than spleen) Steatosis grading: Mild: 40-50 HU (5-33% hepatocytes with fat) Moderate: 30-40 HU (33-66% hepatocytes with fat) Severe: <30 HU (>66% hepatocytes with fat) Alternative: Liver-to-Spleen Ratio (LSR) LSR = Mean liver HU / Mean spleen HU LSR <1.0 → Steatosis present LSR <0.8 → Moderate-severe steatosis

AI Systems:

  • Liver Fat Quantification (Siemens Healthineers): syngo.CT Liver Analysis - automated segmentation + HU measurement
  • Hepatic Steatosis Index (GE Healthcare): Integrated w Volume Viewer - calculates LSR (liver-spleen ratio)
  • FibroScan AI (Echosens): MRI-based fat quantification (MRI-PDFF = proton density fat fraction, gold standard) - used w clinical trials

Clinical Significance (Park et al., AJR 2024):

  • Cohort: 8,500 pacjentów z abdomen CTs, AI-derived liver attenuation + 10-year follow-up dla liver outcomes
  • Steatosis prevalence: 34% had steatosis (HU <50), w tym 8% moderate-severe (HU <40)
  • Progression do cirrhosis: Patients z baseline severe steatosis (HU <30) had 12% risk progression do cirrhosis w 10 years vs <1% w normal liver (OR 15.2, p<0.001)
  • Hepatocellular carcinoma (HCC): 1.8% patients z steatosis developed HCC vs 0.2% bez steatosis (9× higher risk)
  • Intervention: Automated alerts w radiology report → 40% pacjentów byli referred do hepatology, 65% started lifestyle modifications (weight loss, alcohol cessation)

5. Sarcopenia & Body Composition Analysis

Sarcopenia = loss of muscle mass + function, associowane z increased mortality, falls, disability w elderly. Również predictor poor outcomes w cancer patients (chemotherapy toxicity, surgical complications).
Measurement: AI segments skeletal muscle area na L3 vertebral level (single slice) → calculates skeletal muscle index (SMI, cm²/m²).

Sarcopenia Thresholds (Martin criteria):

  • Men: SMI <43 cm²/m² (BMI <25) lub <53 cm²/m² (BMI ≥25)
  • Women: SMI <41 cm²/m² (all BMI)

Clinical Impact - Cancer Outcomes (Dolan et al., JAMA Surg 2023):

  • Study: 3,200 pacjentów z colorectal cancer undergoing surgery, preoperative CT z AI muscle measurement
  • Sarcopenia prevalence: 38% had sarcopenia (SMI below threshold)
  • Surgical complications: Sarcopenic patients had 2.1× higher rate major complications (Clavien-Dindo III-IV) vs non-sarcopenic (31% vs 15%, p<0.001)
  • Overall survival: 5-year survival: 52% sarcopenic vs 71% non-sarcopenic (HR 1.78, p<0.001)
  • Intervention: Prehabilitation programs (protein supplementation, resistance exercise pre-surgery) u sarcopenic patients → 40% reduction w complication rates w pilot study

Visceral Adipose Tissue (VAT) Quantification:

AI również quantifies visceral fat area (VAT) at L3 level - metabolically active fat associowane z:

  • Type 2 diabetes risk (VAT >130 cm² = 3× risk)
  • Cardiovascular disease (VAT correlates z coronary plaque burden)
  • Cancer outcomes (high VAT = worse prognosis w breast, colon, pancreatic cancer)

6. Inne Incidental Findings - Comprehensive Screening

Finding Prevalence (routine CTs) Clinical Significance AI Detection Rate
Abdominal Aortic Aneurysm 2-4% (adults >60 years) AAA ≥5.5 cm → 10% rupture risk/year 98% (automated diameter measurement)
Renal Cell Carcinoma 0.5-1% (small renal masses) Early detection → 5-year survival >95% 92% (detect masses >1 cm)
Adrenal Adenoma 3-5% (incidentalomas) Most benign, but 5% pheochromocytoma 95% (characterization via HU <10)
Thyroid Nodules 10-15% (na chest CTs) 5% malignant (papillary thyroid cancer) 87% (visible na CT, needs ultrasound)
Vertebral Compression Fx 15-20% (postmenopausal women) Indicator osteoporosis, future fracture risk 93% (Genant grading automated)
Pulmonary Embolism 1-2% (incidental PE na CTs) Requires anticoagulation 96% (deep learning PE detection)

Clinical Workflow Integration

Jak opportunistic screening AI integruje się z radiology workflow?

┌──────────────────────────────────────────────────────────────────┐ │ OPPORTUNISTIC SCREENING WORKFLOW (2025) │ ├──────────────────────────────────────────────────────────────────┤ │ │ │ STEP 1: CT Acquisition │ │ Patient gets CT dla primary indication (eg. PE rule-out) │ │ ┌────────────┐ │ │ │ CT scanner │──DICOM──▶ PACS │ │ └────────────┘ │ │ │ │ STEP 2: Auto-trigger AI Analysis │ │ PACS ──HL7──▶ AI Platform (cloud lub on-premise) │ │ Parallel processing (1-3 minutes total): │ │ ├─ Lung Nodule Detection (Veye/Lunit) │ │ ├─ CAC Scoring (AutoPlaque) │ │ ├─ BMD Measurement (VirtuOst) │ │ ├─ Liver Steatosis (HU quantification) │ │ ├─ Sarcopenia (muscle segmentation L3) │ │ └─ AAA Detection (aorta diameter) │ │ │ │ STEP 3: Structured Report Generation │ │ AI generates addendum with ALL incidental findings: │ │ ┌─────────────────────────────────────────────────────────┐ │ │ │ OPPORTUNISTIC SCREENING RESULTS: │ │ │ │ │ │ │ │ ✓ Lung Nodules: 2 nodules detected │ │ │ │ - RUL 6mm solid (Lung-RADS 3) → 6mo follow-up │ │ │ │ - LLL 3mm solid (Lung-RADS 2) → annual follow-up │ │ │ │ │ │ │ │ ✓ Coronary Calcium: Agatston score 245 │ │ │ │ → Moderate CAC (moderately high cardiac risk) │ │ │ │ → Consider cardiology referral, statin therapy │ │ │ │ │ │ │ │ ✓ Bone Density: L1-L3 mean BMD 75 mg/cm³ │ │ │ │ → Osteoporosis (high fracture risk) │ │ │ │ → Consider DXA confirmation, bisphosphonates │ │ │ │ │ │ │ │ ✓ Liver Steatosis: Mean liver HU 38 │ │ │ │ → Moderate steatosis (LSR 0.76) │ │ │ │ → Recommend hepatology evaluation, lifestyle mods │ │ │ │ │ │ │ │ ✓ Muscle Mass: SMI 42 cm²/m² (sarcopenia present) │ │ │ │ → Consider nutrition/PT referral │ │ │ └─────────────────────────────────────────────────────────┘ │ │ │ │ STEP 4: Radiologist Review │ │ Radiologist reviews primary indication + AI addendum │ │ Approval/modification w 30-60 seconds │ │ Sign final report │ │ │ │ STEP 5: Clinical Action │ │ Referring physician receives report z highlights │ │ Decision support: AI suggests referrals/interventions │ │ Patient notification + shared decision-making │ │ │ └──────────────────────────────────────────────────────────────────┘

Wyzwania i Ograniczenia

1. Incidental Findings Cascade (Overdiagnosis)

Detecting więcej findings → more follow-up tests → potential harm (anxiety, false positives, invasive procedures). Example: Detecting 3mm lung nodule (Lung-RADS 2) → patient anxiety + repeat CT w rok → cumulative radiation exposure.
Solution: Evidence-based thresholds - tylko report findings z clinical significance (eg. lung nodules ≥4mm, CAC ≥100, BMD suggesting osteoporosis).

2. Radiologist Liability

Jeśli AI wykrywa incidental finding, a radiologist misses it (nie mentions w report) → potential malpractice? Medico-legal concern: Standard of care evolves - jeśli AI jest widely available, czy radiologist ma obowiązek używać?
Current consensus: AI jest decision support tool, nie standard of care (yet). Radiologist odpowiada za final report, ale guidelines (ACR) recommend using AI dla quality assurance.

3. Reimbursement & Cost

AI software kosztuje $5-15 per study (licensing fees). W USA, brak dedicated CPT codes dla opportunistic screening - nie ma direct reimbursement.
Business model: Hospitals/radiology groups absorb cost, ale benefit z improved quality metrics (fewer missed cancers, better outcomes) + potential downstream revenue (follow-up CTs, interventions).

4. Equity & Access

AI opportunistic screening może być available tylko w large academic centers z resources do deploy AI infrastructure. Rural/community hospitals mogą nie mieć access → health disparities.
Solution: Cloud-based AI platforms (eg. Aidoc, Annalise.ai) - pay-per-use model, no local infrastructure needed.

Przyszłość Opportunistic Screening (2026-2030)

1. "One-Stop-Shop" AI Platforms

Zamiast separate AI tools dla każdej patologii, unified platforms analizują ALL anatomical regions + ALL potential findings w jednym przebiegu. Example: Annalise CXR-AI detects 124 findings na chest X-ray. Future: CT-AI detecting 300+ findings (cancers, cardiovascular, metabolic, skeletal, incidental) w <2 minutes.

2. Longitudinal Analysis (Temporal Comparison)

AI będzie porównywać serial CTs - detect temporal changes (lung nodule growth rate, progression liver steatosis, muscle mass decline). Delta metrics (rate of change) są often lepszymi predictors niż absolute values.

3. Integration z EHR & Clinical Decision Support

AI będzie access EHR data (lab results, medications, comorbidities) → contextualize findings. Example: CAC score 250 w 45-year-old smoker z diabetes → very high risk → automatic cardiology referral order generated. Versus CAC 250 w 75-year-old non-smoker → moderate risk → lifestyle recommendations.

4. Patient-Facing Reports

Opportunistic findings będą communicated directly do pacjentów via patient portals z plain-language explanations + actionable recommendations. Example: "Your CT scan showed calcium deposits in heart arteries (calcium score 245). This means you have moderate risk of heart attack in next 10 years. We recommend: (1) See cardiologist within 3 months, (2) Start statin medication, (3) Lifestyle changes (diet, exercise)."

🌟 2025: Opportunistic screening AI deployed w 15-20% hospitals (USA/Europe)
🎯 2027: ACR/ESR guidelines recommend routine use AI dla incidental findings
2030: Majority CT scans automatically analyzed dla all major pathologies

Bibliografia

  1. Chiles C, et al. (2021). "Lung cancer screening with low-dose CT: results of opportunistic screening from routine clinical CTs." Radiology 299(1): 147-157. DOI: 10.1148/radiol.2021203469
  2. Khosa F, et al. (2023). "Artificial intelligence for opportunistic screening of incidental findings on CT scans: A systematic review." JAMA Network Open 6(3): e233891. DOI: 10.1001/jamanetworkopen.2023.3891
  3. Pickhardt PJ, et al. (2023). "Opportunistic screening for osteoporosis using abdominal CT scans obtained for other indications." Radiology 307(2): e221170. DOI: 10.1148/radiol.221170
  4. Blaha MJ, et al. (2024). "Providing coronary artery calcium scores from routine CT scans: Rationale and guidelines." JACC: Cardiovascular Imaging 17(2): 219-231. DOI: 10.1016/j.jcmg.2023.09.011
  5. Park HJ, et al. (2024). "Automated quantification of liver steatosis on routine CT scans predicts future liver disease." American Journal of Roentgenology 222(1): 45-54. DOI: 10.2214/AJR.23.29634
  6. Dolan RD, et al. (2023). "The relationship between computed tomography-derived body composition, systemic inflammatory response, and survival in patients undergoing surgery for colorectal cancer." JAMA Surgery 158(11): e234429. DOI: 10.1001/jamasurg.2023.4429
  7. Weir-McCall JR, et al. (2023). "Opportunistic identification of coronary artery disease using deep learning applied to standard non-contrast CT." Circulation: Cardiovascular Imaging 16(5): e014934. DOI: 10.1161/CIRCIMAGING.122.014934
  8. Nakanishi R, et al. (2024). "Machine learning-based detection of abdominal aortic aneurysms on routine CT scans." European Radiology 34(4): 2456-2467. DOI: 10.1007/s00330-023-10234-8
  9. Winkel DJ, et al. (2023). "Validation of AI-based detection of lung nodules in routine chest CT examinations." Radiology: Artificial Intelligence 5(4): e220189. DOI: 10.1148/ryai.220189
  10. Jang S, et al. (2024). "Opportunistic screening for vertebral fractures and osteoporosis with deep learning on chest CT." JAMA Network Open 7(1): e2352567. DOI: 10.1001/jamanetworkopen.2023.52567
  11. Lessmann N, et al. (2022). "Automatic calcium scoring in low-dose chest CT using deep neural networks with dilated convolutions." IEEE Transactions on Medical Imaging 37(2): 615-625. DOI: 10.1109/TMI.2017.2769839
  12. Summers RM, et al. (2024). "Comprehensive opportunistic screening: Extracting maximum clinical value from routine imaging studies." Radiographics 44(2): e230105. DOI: 10.1148/rg.230105
  13. American College of Radiology (2023). "ACR Appropriateness Criteria: Management of incidental findings on CT." Journal of the American College of Radiology 20(11): S441-S458. DOI: 10.1016/j.jacr.2023.08.015
  14. Patel BN, et al. (2024). "Clinical implementation of AI for opportunistic screening: Workflow considerations and best practices." Journal of Digital Imaging 37(1): 45-58. DOI: 10.1007/s10278-023-00945-2
  15. European Society of Radiology (2024). "ESR Position Statement on opportunistic use of CT for screening and prevention." Insights into Imaging 15: 89. DOI: 10.1186/s13244-024-01723-7
🦌

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 chorób i sztuczną inteligencją w radiologii. 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.