Interpreting AI Results

Understanding Neolens AI outputs is crucial for safe and effective clinical integration. This guide helps you interpret confidence scores, classifications, and recommendations appropriately.

---

📊 Understanding Confidence Scores

Confidence Levels Explained

Score Range	Interpretation	Clinical Action	Visual Indicator
0.90 - 1.00	Very High Confidence	Priority review recommended	🔴 Red highlight
0.75 - 0.89	High Confidence	Standard workflow	🟡 Orange highlight
0.60 - 0.74	Moderate Confidence	Consider additional imaging	🟡 Yellow highlight
0.40 - 0.59	Low Confidence	Human review essential	⚪ Gray highlight
0.00 - 0.39	Very Low Confidence	Likely false positive	No highlight

Example Response Structure

{
  "findings": [
    {
      "label": "pulmonary_nodule",
      "confidence": 0.87,
      "location": {
        "x": 245, "y": 156,
        "width": 32, "height": 28
      },
      "clinical_significance": "moderate",
      "follow_up_recommended": true
    }
  ],
  "overall_assessment": {
    "normal_probability": 0.23,
    "abnormal_probability": 0.77,
    "urgency": "routine"
  }
}

---

🏷️ Classification Categories

Pathology Labels

Neolens uses standardized medical terminology aligned with:

• ICD-11 diagnostic codes
• RadLex radiology lexicon
• SNOMED CT clinical terminology

Common Classifications

Thoracic Imaging:

• pneumothorax - Collapsed lung
• pleural_effusion - Fluid in pleural space
• pulmonary_nodule - Lung nodule
• cardiomegaly - Enlarged heart
• atelectasis - Lung collapse

Neuroimaging:

• hemorrhage - Brain bleeding
• ischemic_stroke - Blood clot-related stroke
• mass_effect - Space-occupying lesion
• hydrocephalus - Fluid accumulation
• midline_shift - Brain structure displacement

Abdominal Imaging:

• bowel_obstruction - Intestinal blockage
• appendicitis - Appendix inflammation
• free_air - Pneumoperitoneum
• hepatomegaly - Enlarged liver

---

📐 Measurement Interpretation

Anatomical Measurements

{
  "measurements": {
    "aortic_diameter": {
      "value": 4.2,
      "unit": "cm",
      "reference_range": "2.0-3.5 cm",
      "status": "enlarged",
      "clinical_significance": "moderate_aneurysm"
    },
    "left_ventricle": {
      "ejection_fraction": 45,
      "unit": "percent",
      "reference_range": "50-70%",
      "status": "reduced",
      "clinical_significance": "mild_dysfunction"
    }
  }
}

Measurement Accuracy

• Linear measurements: ±2-3mm accuracy
• Area calculations: ±5-8% variability
• Volume estimates: ±10-15% uncertainty
• Angles: ±2-5 degrees precision

---

🚨 Priority and Urgency Indicators

Urgency Classifications

Level	Description	Response Time	Examples
critical	Life-threatening findings	< 1 hour	Hemorrhage, pneumothorax
urgent	Requires prompt attention	< 4 hours	Large masses, fractures
routine	Standard follow-up	< 24 hours	Small nodules, mild changes
incidental	Unexpected findings	Variable	Benign cysts, artifacts

Priority Scoring Algorithm

Priority Score = (Confidence × 0.4) + (Clinical Significance × 0.3) + (Size/Extent × 0.3)

Where:
- Confidence: 0.0-1.0
- Clinical Significance: 0.0-1.0 (based on pathology severity)
- Size/Extent: 0.0-1.0 (normalized to anatomy)

---

🔄 Contextualizing Results

Patient History Integration

Consider AI results alongside:

• Demographics: Age, gender, medical history
• Clinical presentation: Symptoms, vital signs
• Prior imaging: Comparison studies, disease progression
• Laboratory data: Biomarkers, blood work

Imaging Context

• Modality limitations: CT vs MRI vs X-ray capabilities
• Acquisition parameters: Slice thickness, contrast timing
• Image quality: Motion artifacts, noise levels
• Technical factors: Scanner type, reconstruction algorithms

---

📈 Longitudinal Analysis

Tracking Changes Over Time

{
  "longitudinal_analysis": {
    "baseline_study": "2024-01-15",
    "current_study": "2024-08-05",
    "changes_detected": [
      {
        "finding": "pulmonary_nodule",
        "baseline_size": "8mm",
        "current_size": "12mm",
        "growth_rate": "4mm over 6 months",
        "doubling_time": "18 months",
        "recommendation": "short_term_follow_up"
      }
    ]
  }
}

Growth Pattern Analysis

• Doubling time calculations
• Volume change percentages
• Shape evolution tracking
• New findings identification

---

⚖️ Uncertainty and Limitations

Understanding AI Uncertainty

{
  "uncertainty_metrics": {
    "epistemic_uncertainty": 0.12,    // Model uncertainty
    "aleatoric_uncertainty": 0.08,    // Data uncertainty
    "total_uncertainty": 0.20,
    "confidence_interval": [0.73, 0.91],
    "prediction_explanation": "High contrast lesion with clear borders"
  }
}

When to Be Cautious

• Edge cases: Rare pathologies, unusual presentations
• Poor image quality: Motion artifacts, low resolution
• Pediatric patients: Limited training data
• Implants/hardware: Potential interference
• Multi-pathology cases: Complex interactions

---

🎯 Clinical Decision Support

Recommendation Types

{
  "recommendations": {
    "imaging": [
      {
        "type": "follow_up_ct",
        "timeframe": "3_months",
        "reason": "monitor_nodule_growth",
        "contrast": "without"
      }
    ],
    "clinical": [
      {
        "action": "pulmonology_referral",
        "urgency": "routine",
        "reason": "nodule_evaluation"
      }
    ],
    "additional_testing": [
      {
        "test": "pet_scan",
        "indication": "characterize_lesion",
        "priority": "consider"
      }
    ]
  }
}

Treatment Pathway Integration

• Guidelines compliance: NCCN, ACR, ESR protocols
• Risk stratification: Low, intermediate, high risk
• Cost-effectiveness: Optimize imaging utilization
• Patient preferences: Shared decision-making support

---

🔍 Visual Interpretation Aids

Heatmaps and Overlays

• Attention maps: Areas of AI focus
• Confidence overlays: Visual confidence representation
• Segmentation masks: Anatomical structure outlines
• Comparison views: Side-by-side analysis

Annotation Features

{
  "annotations": {
    "bounding_boxes": [
      {
        "coordinates": [120, 80, 240, 160],
        "label": "suspicious_area",
        "confidence": 0.85
      }
    ],
    "segmentation_masks": {
      "lung_fields": "base64_encoded_mask",
      "heart_contour": "base64_encoded_mask"
    },
    "arrows_and_callouts": [
      {
        "type": "arrow",
        "start": [150, 120],
        "end": [180, 140],
        "label": "area_of_concern"
      }
    ]
  }
}

---

⚠️ Common Interpretation Pitfalls

False Positives

• Artifacts mistaken for pathology
• Normal variants flagged as abnormal
• Overlapping structures misidentified

False Negatives

• Subtle findings missed
• Atypical presentations
• Poor image quality masking pathology

Overreliance on AI

• Confirmation bias - Accepting AI results without scrutiny
• Automation bias - Reduced human vigilance
• Context neglect - Ignoring clinical picture

---

📚 Best Practices for Result Interpretation

✅ Do

• Always correlate with clinical context
• Review original images, not just AI overlays
• Consider differential diagnoses
• Document AI assistance in reports
• Maintain clinical reasoning skills

❌ Don't

• Accept results without clinical correlation
• Ignore low-confidence findings completely
• Skip verification of critical findings
• Rely solely on AI for urgent cases
• Forget to consider image quality issues

---

🔗 Integration with Clinical Workflow

PACS Integration

<!-- DICOM Structured Report Example -->
<DicomSR>
  <Finding>
    <Text>AI-detected pulmonary nodule</Text>
    <Confidence>0.87</Confidence>
    <Location>RUL posterior segment</Location>
    <Measurements>8mm diameter</Measurements>
  </Finding>
</DicomSR>

Report Generation

Structured reporting templates integrate AI findings:

• Impression section: AI-highlighted abnormalities
• Recommendations: Evidence-based follow-up
• Comparison: Prior study analysis
• Limitations: AI uncertainty documentation

---

Key Takeaways

• Confidence scores guide clinical attention but don't replace judgment
• Always consider clinical context and image quality
• Use AI as a diagnostic aid, not replacement for expertise
• Document AI assistance for regulatory compliance
• Maintain skeptical evaluation of all AI outputs

---

🔗 Related Documentation

• Configuring AI
• AI Limitations
• Best Practices
• Uncertainty Estimation