Note for Pathologists: A summary of quality issues encountered during Phase 1 that negatively impacted AI performance. Addressing these is crucial for reliable AI deployment.
### Quality Issues from Phase 1:From the first phase analysis, the following quality-related issues were identified:1. **Blurred Images**: 3 images were too blurred for reliable AI analysis2. **Processing Artifacts**: 4 cores had processing artifacts that affected AI performance3. **Focus Issues**: 1 case had both blurred and focused slides4. **Total Excluded**: 7 cores were excluded from agreement analysis due to quality issuesPre-Analysis Quality Assessment
Based on Phase 1 findings, implementing quality control before AI analysis is critical.
Note for Pathologists: A proposed Quality Control workflow outlining three phases: retrospective assessment, real-time prospective QC, and continuous monitoring.
**1. Automated Quality Assessment**- Use HistoQC (https://github.com/histopathology/HistoQC) for automated quality metrics- Assess focus, artifacts, tissue coverage, and staining quality- Flag images below quality thresholds before AI processing**2. Quality Metrics to Monitor**- Focus score (detect blurred regions)- Artifact detection (bubbles, folds, processing issues)- Tissue coverage (adequate tissue present)- Staining intensity and uniformity- Background clarity**3. Quality Gates**- Define minimum acceptable quality thresholds- Require manual review for images below threshold- Re-scan poor quality slides when possible- Document quality issues in diagnostic workflowFalse Positive/Negative Analysis Related to Quality
Note for Pathologists: Analysis of False Negatives from Phase 1 reveals that the majority were caused by suboptimal image quality, highlighting the direct link between slide quality and AI sensitivity.
### False Negatives (AI missed tumor):From Phase 1, 8 cores were diagnosed as benign by AI but malignant by pathologists:- 3 images: Blurred (quality issue)- 4 images: Processing artifacts (quality issue)- 1 image: True false negative (confirmed with IHC)**Quality-related false negatives: 7/8 (87.5%)**### Quality Impact Summary:- Quality issues accounted for majority of AI false negatives- This suggests AI performance is highly dependent on image quality- Implementing QC could significantly reduce false negative rateNote for Pathologists: Proposed quantitative metrics and thresholds for automated quality acceptance.
| Metric | Threshold | Description | Action_if_Failed |
|---|---|---|---|
| Focus Score | > 0.7 | Measure of image sharpness (0-1 scale) | Flag for review / Re-scan |
| Artifact Score | < 0.2 | Percentage of image with artifacts (0-1 scale) | Flag for review / Manual diagnosis |
| Tissue Coverage | > 60% | Percentage of image containing tissue | May be acceptable if diagnostic area present |
| Staining Quality | 0.6 - 0.9 | Staining intensity uniformity | Flag for review / Check staining protocol |
| Background Score | > 0.8 | Clear background without debris | Review if combined with other issues |
Assess Current Image Quality
**Steps:**1. Run HistoQC on all existing digital slides2. Generate quality metrics for entire archive3. Identify systematic quality issues4. Determine appropriate thresholds for institution5. Correlate quality metrics with AI performanceImplement Real-Time QC
**Implementation:**1. Integrate HistoQC into scanning workflow2. Automatic quality assessment post-scanning3. Quality dashboard for lab technologists4. Alert system for below-threshold images5. Track quality metrics over time6. Only send quality-approved images to AI systemOngoing Quality Assurance
**Monitoring Plan:**1. Weekly quality metric reports2. Identify trends in quality issues3. Scanner maintenance based on quality decline4. Operator training for common issues5. Quarterly review of quality thresholds6. Document correlation with AI performanceNote for Pathologists: A checklist detailing QC tasks at each stage of the digital pathology workflow, assigning responsibilities to ensuring high-quality input for the AI.
| Stage | Task | Responsible |
|---|---|---|
| Pre-Scanning | Verify slide quality under microscope | Histotechnologist |
| Pre-Scanning | Check staining quality | Histotechnologist |
| Pre-Scanning | Ensure proper tissue mounting | Histotechnologist |
| Post-Scanning | Run automated quality assessment | Automated System |
| Post-Scanning | Review quality metrics | Lab Technologist |
| Post-Scanning | Flag problematic images | Lab Technologist |
| Pre-AI Analysis | Verify image passed QC thresholds | System / Pathologist |
| Pre-AI Analysis | Manual review of flagged images | Pathologist |
| Post-AI Analysis | Track AI performance on QC-passed images | System Administrator |
| Post-AI Analysis | Document quality-related issues | Pathologist / Admin |
Note for Pathologists: A cost-benefit breakdown showing that the minimal investment in QC software and training is outweighed by significant clinical benefits.
**Benefits of Implementing QC:**1. **Reduced False Negatives**: Phase 1 showed 87.5% of AI false negatives were quality-related2. **Improved AI Accuracy**: Only analyze high-quality images3. **Better Pathologist Confidence**: Knowing images meet quality standards4. **Reduced Re-work**: Identify issues before diagnosis5. **Scanner Maintenance**: Early detection of scanner problems**Costs of Implementing QC:**1. **Software**: HistoQC is open-source (free)2. **Computation**: Additional processing time per image (~1-2 minutes)3. **Storage**: Quality metrics and reports (~1MB per case)4. **Training**: Staff training on QC workflow (~4 hours)5. **Monitoring**: Weekly review time (~1 hour/week)**Net Benefit**: The quality control costs are minimal compared to the benefit of improved diagnostic accuracy and reduced AI errors.Using HistoQC for Automated Quality Assessment
HistoQC is an open-source quality control system for digital pathology that can automatically assess image quality.
### HistoQC Features:- Automated detection of common artifacts- Focus quality assessment- Tissue detection and coverage- Staining quality evaluation- Background and pen marking detection- Generates quality reports and visualizations### Example HistoQC Integration:```bash# Run HistoQC on a directory of imagespython -m histoqc.pipeline --config config.ini --nprocesses 4 slides/*.svs```### Configuration Recommendations:1. **Enable Modules**: - Brightness/Contrast - Blur Detection - Bubble Detection - Fold Detection - Tissue Coverage2. **Set Thresholds**: Based on institution-specific requirements3. **Generate Reports**: HTML reports with quality visualizations4. **Integration Points**: - Post-scanning automated trigger - Quality database for tracking - Dashboard for lab monitoring - API for AI system integrationKey Recommendations:
Expected Outcomes: - Reduced AI false negatives (potentially 87.5% reduction based on Phase 1) - Improved diagnostic confidence - Earlier detection of technical issues - Better overall workflow efficiency
Critical Success Factor: Quality control must be implemented BEFORE AI analysis, not after.
This chapter analyzes the impact of image quality on AI performance. Phase 1 results indicated that processing artifacts and image quality issues significantly affected AI accuracy.