Missed Lung Cancer on CT: AI Detection, Diagnostic Delay, and Economic Impact

- 9월 18, 2026

Edited by ScholarGen TheMediTech Team

Executive Summary

A pulmonary nodule that is overlooked on CT may appear insignificant at the time of the original examination. Six months later, however, the same lesion may be larger, morphologically suspicious, and clinically consequential.

This is one of the most important challenges in modern thoracic imaging.

The problem is not simply that a radiologist “missed a cancer.” In many real-world situations, diagnostic delay can emerge from several interconnected failures:

subtle lesion detection,
perceptual oversight,
comparison with prior examinations,
reporting,
follow-up recommendation,
communication,
and longitudinal tracking.

The clinical importance of early detection is substantial. Global cancer statistics identify lung cancer as the leading cause of cancer death worldwide, with approximately 1.8 million deaths in 2022. In contemporary U.S. SEER data, five-year relative survival is substantially higher when lung cancer is localized than when it is diagnosed after distant spread.

This creates a practical question for healthcare organizations:

Can artificial intelligence reduce the probability that a clinically meaningful pulmonary nodule will be overlooked, and can an integrated AI workflow generate measurable economic value?

The answer requires more nuance than simply purchasing an AI algorithm.

AI-assisted pulmonary nodule detection can highlight candidate lesions, support measurements, compare current and prior examinations, and help prioritize cases. Deep-learning research has demonstrated clinically relevant performance in lung cancer screening and pulmonary nodule detection.

However, AI does not eliminate diagnostic error.

The highest-value implementation is therefore not:

AI → radiologist

but rather:

CT → AI detection → radiologist review → longitudinal comparison → structured reporting → follow-up tracking → clinical action

The economic benefit should also be evaluated as a healthcare-system model rather than as a guaranteed dollar amount. Earlier-stage diagnosis may reduce downstream treatment burden in some patients, but the actual financial effect varies according to stage, histology, treatment pathway, payer, geography, and institutional costs.

1. Why Are Small Lung Cancers Missed on CT?

Pulmonary nodules can be difficult to detect even on high-quality CT.

The challenge becomes particularly important when a lesion is:

only a few millimeters in diameter,
located adjacent to vessels,
close to the pleura,
partially obscured by atelectatic lung,
associated with motion artifact,
or visually similar to normal anatomical structures.

A busy reading environment adds another layer of risk.

Radiologists do not inspect every pixel with equal attention. Interpretation is a highly complex perceptual and cognitive task involving search, pattern recognition, comparison, prioritization, and clinical reasoning.

A subtle nodule can therefore exist on an image without becoming sufficiently salient to the reader.

Table 1. Common Mechanisms of Pulmonary Nodule Misses

Failure mechanism	Imaging example	Potential consequence
Small lesion size	3–5 mm nodule	Low visual conspicuity
Vessel overlap	Nodule adjacent to pulmonary vessel	Vessel–nodule confusion
Partial-volume averaging	Thick reconstructed section	Reduced lesion conspicuity
Motion artifact	Lower-lobe blurring	Boundary obscuration
Similar attenuation	Ground-glass or faint solid nodule	Low contrast
Multiple abnormalities	Numerous incidental findings	Search fatigue
Prior-study mismatch	No efficient comparison	Growth not recognized
Workflow overload	High-volume reading session	Perceptual/cognitive error
Reporting omission	Lesion not included in final report	No clinical follow-up
Follow-up failure	Recommendation not tracked	Delayed diagnosis

The technical component is also important.

The Fleischner Society recommends thin-section chest CT reconstruction for accurate characterization and measurement of small nodules; thick sections can increase volume averaging and interfere with characterization.

2. The Clinical Importance of Early Detection

The biological behavior of lung cancer varies substantially among patients and tumor subtypes, but stage at diagnosis remains a major determinant of outcome.

Current U.S. SEER statistics show a marked difference between localized and distant lung and bronchus cancer:

SEER stage	5-year relative survival
Localized	65.5%
Regional	38.2%
Distant	10.5%
All stages	29.5%

These figures are population-level statistics rather than predictions for an individual patient.

The American Cancer Society reports similarly large stage-associated differences for NSCLC, with approximately 67% five-year relative survival for localized disease compared with 12% for distant disease in its cited SEER dataset.

This is why the clinical objective of AI-assisted CT analysis should not simply be:

“Find more nodules.”

The more meaningful objective is:

Identify clinically relevant abnormalities early enough to change the patient's diagnostic pathway.

3. CT Physics: Why a 4-mm Nodule Can Be Difficult to See

A pulmonary nodule is not simply an object that either exists or does not exist.

Its visibility depends on:

spatial resolution,
contrast resolution,
reconstruction thickness,
reconstruction kernel,
noise,
lesion-to-lung attenuation difference,
motion,
vessel overlap,
and the observer's visual attention.

Partial-Volume Effect

When a small nodule occupies only part of a CT voxel, its measured attenuation is averaged with surrounding lung.

Consequently, a small solid lesion may appear less dense than its true physical attenuation.

This is one reason thin-section reconstruction is important.

The Fleischner Society recommends contiguous thin sections of approximately 1.0–1.5 mm for thoracic CT interpretation and nodule characterization.

Figure 1. Conceptual Effect of Section Thickness

Clinical message: A technically adequate CT acquisition and reconstruction is part of the diagnostic system. AI cannot fully compensate for inadequate image quality.

4. A Representative Missed-Nodule Scenario

Consider a representative clinical scenario.

A patient undergoes chest CT.

Initial Examination

Right upper-lobe pulmonary nodule
Approximately 4 mm
Subtle morphology
Adjacent to vascular structures
No specific mention in the final report

At this point, the lesion may be below the threshold at which its clinical significance is obvious.

Six months later, another CT is obtained.

Follow-up Examination

The lesion now measures approximately 9 mm.

Its morphology has become more conspicuous:

increased diameter,
irregular contour,
greater soft-tissue conspicuity,
suspicious interval growth.

A diagnostic work-up is initiated, eventually leading to a diagnosis of lung cancer.

Table 2. Longitudinal Change in the Representative Scenario

Parameter	Initial CT	Follow-up CT
Approximate diameter	4 mm	9 mm
Visibility	Subtle	Conspicuous
Margin	Difficult to characterize	Irregular
Clinical interpretation	Not reported	Suspicious
Management	No documented pathway	Diagnostic work-up
Potential stage impact	Earlier disease potentially possible	Potentially more advanced disease

Important: This is a representative clinical scenario for workflow analysis, not a claim that every 4-mm nodule represents cancer or that a six-month interval necessarily causes stage migration.

Management of small nodules depends on morphology, patient risk, size, and whether the examination is performed for screening or incidental detection. The Fleischner recommendations explicitly emphasize individualized management.

5. The Real Problem: Three Separate Failures

A missed pulmonary nodule should not automatically be interpreted as a single-person error.

There may be at least three separate failure points.

Failure 1 — Detection

The lesion is present but not recognized.

Failure 2 — Reporting

The lesion is recognized but is not adequately documented or communicated.

Failure 3 — Follow-up

The lesion is reported, but the recommended follow-up does not occur.

These are fundamentally different problems.

Figure 2. The Diagnostic Failure Chain

The most important insight is that AI detection alone cannot solve the entire chain.

6. What AI Can Actually Do

Modern pulmonary AI systems can support several tasks.

6.1 Candidate Detection

The algorithm scans the CT volume and identifies candidate pulmonary nodules.

6.2 Localization

The candidate is marked directly on the CT series.

6.3 Measurement

The system may provide:

maximum diameter,
volumetric measurement,
three-dimensional segmentation,
density characteristics.

6.4 Prior Examination Comparison

A more sophisticated system can compare current and previous CT examinations.

6.5 Growth Analysis

Instead of asking only:

“Is there a nodule?”

the system can ask:

“Has this nodule changed?”

This is clinically important because interval change is often more informative than a single measurement.

6.6 Risk Stratification

Some systems can combine imaging features with patient information to estimate malignancy risk.

Deep-learning research has demonstrated that algorithms using current and prior CT volumes can perform lung cancer risk prediction, illustrating the potential value of longitudinal rather than single-time-point analysis.

7. AI Should Be a Second Reader, Not an Autonomous Judge

8. Why AI Tracking May Matter More Than AI Detection

Detection is only the beginning.

Imagine two systems.

System A

“There is a 6-mm nodule.”

System B

“There is a 6-mm nodule that was 4 mm six months ago and has increased in volume.”

The second output is potentially more clinically useful.

This is the transition from:

AI detection

to:

AI longitudinal disease surveillance.

Table 3. Detection vs. Tracking

Capability	Detection AI	Longitudinal AI
Find nodule	✓	✓
Localize lesion	✓	✓
Measure size	✓	✓
Compare prior CT	Limited	✓
Growth calculation	Limited	✓
Follow-up reminder	Usually no	Possible
Longitudinal risk	Limited	Potentially higher
Enterprise integration	Optional	Highly valuable

The long-term opportunity therefore lies in connecting image interpretation with longitudinal patient management.

9. Enterprise AI Workflow

10. PACS Integration Is Not Enough

A common implementation mistake is to stop at PACS.

An AI result appearing inside the radiologist's workstation is useful, but it does not guarantee that the patient will receive appropriate follow-up.

A complete system should ideally connect:

PACS → RIS → reporting → EHR → follow-up management

The final step is critical.

A radiologist may correctly write:

“Recommend follow-up CT in 6 months.”

But if nobody monitors whether the patient actually returns, the clinical workflow remains incomplete.

11. Clinical Decision Tree

Figure 3. AI-Assisted Pulmonary Nodule Decision Framework

This framework is deliberately simplified. Actual management must follow the applicable clinical guideline and patient-specific risk assessment.

12. Differential Diagnosis

A pulmonary nodule is not synonymous with lung cancer.

The differential diagnosis includes:

Diagnosis	Typical imaging clue
Granuloma	Calcification, characteristic morphology
Hamartoma	Fat or characteristic “popcorn” calcification
Metastasis	Multiple nodules or known primary malignancy
Infection	Clinical context, inflammatory features
Scar	Linear/pleural-associated morphology
Primary lung cancer	Suspicious morphology and/or interval growth
Benign intrapulmonary lymph node	Typical perifissural location and morphology

Therefore, an AI system that simply increases the number of detected nodules may also increase false-positive workload.

The goal is not:

maximum detection

but:

clinically useful detection.

13. The False-Positive Problem

Every additional AI alert has a cost.

A false-positive candidate may lead to:

additional image review,
repeat CT,
specialist consultation,
patient anxiety,
unnecessary procedures,
increased radiologist workload.

Therefore, an AI implementation should monitor both:

Sensitivity

How many clinically important lesions are detected?

Specificity / False-Positive Burden

How many unnecessary alerts are generated?

This is why pulmonary AI performance should not be summarized by a single sensitivity number.

A hospital should evaluate the algorithm in its own population and workflow.

14. Economic Impact: Where Can the Savings Come From?

The original assumption that early-stage lung cancer costs “$20,000–50,000” while late-stage disease costs “$100,000+” is too broad to apply universally.

Actual costs vary according to:

country,
payer,
treatment modality,
cancer stage,
histology,
molecular profile,
surgery,
radiation,
systemic therapy,
hospitalization,
complications,
and survival duration.

For example, economic research in South Korea found that lung cancer treatment costs varied substantially according to treatment strategy, with patients receiving surgery having different lifetime treatment expenditures from those receiving chemotherapy/radiotherapy or supportive treatment.

Therefore, a hospital should use a scenario-based economic model.

Table 4. Illustrative Economic Model

Economic component	Without integrated AI	With integrated AI
Missed-nodule probability	Baseline	Potentially reduced
Additional diagnostic delay	Possible	Potentially reduced
Repeat diagnostic imaging	Variable	Potentially reduced
Advanced-stage treatment	Potentially higher	Potentially lower in selected cases
Follow-up loss	Possible	Trackable
Radiologist workload	Baseline	May shift toward verification
AI licensing	None	Added
Integration cost	None	Added
IT maintenance	Baseline	Added
Potential avoided cost	—	Case-dependent

15. Can AI Really Save $100,000?

It can be possible in a particular case, but it should not be presented as a guaranteed saving.

A more defensible economic equation is:

Net Economic Value = Avoided downstream cost − AI licensing − integration − maintenance − additional diagnostic costs

For example, consider a hypothetical hospital:

Annual AI licensing: $50,000
Integration and implementation: $30,000
Maintenance and governance: $20,000
Total first-year technology expenditure: $100,000

Suppose the system contributes to earlier recognition of several clinically meaningful cancers.

If the institution can document that avoided downstream costs exceed the technology investment, the program may demonstrate positive economic value.

But that conclusion must be derived from actual institutional data.

Figure 4. ROI Sensitivity Model

The most important variable is therefore not the price of the algorithm.

It is:

How many clinically meaningful diagnostic delays does the system actually prevent?

16. ROI Should Be Measured at the Hospital Level

A mature AI program should track operational and clinical metrics simultaneously.

Table 5. Hospital AI ROI Dashboard

Domain	Suggested KPI
Detection	Sensitivity for clinically significant nodules
False positives	AI alerts per 100 CT examinations
Workflow	Median radiologist review time
Follow-up	Completion rate of recommended CT
Diagnosis	Time from CT to definitive diagnosis
Stage	Stage distribution at diagnosis
Safety	AI-related adverse events
Economics	Cost per additional cancer detected
Finance	Avoided downstream expenditure
User experience	Radiologist acceptance
Reliability	Algorithm uptime
Governance	Version/model-drift monitoring

This converts AI from a technology purchase into a measurable clinical program.

17. Regulatory Perspective: What Does FDA 510(k) Clearance Mean?

FDA 510(k) clearance should not be interpreted as a universal ranking of diagnostic accuracy.

The 510(k) pathway evaluates whether a device is substantially equivalent to a legally marketed predicate device, with review of relevant safety and performance information.

FDA explicitly describes 510(k) as a comparative regulatory pathway.

Therefore:

510(k) clearance ≠ “the most accurate AI.”

It also does not mean that a cleared algorithm will perform identically in every hospital.

Real-world performance can change with:

scanner vendors,
reconstruction parameters,
patient population,
prevalence,
workflow,
radiologist behavior,
and software version.

18. Explainable AI in Pulmonary Imaging

A clinically useful AI system should show the radiologist why a case has been flagged.

Useful visualization may include:

nodule bounding box,
segmentation contour,
diameter,
volume,
previous measurement,
growth percentage,
confidence score,
and comparison images.

Figure 5. Ideal AI Result

The purpose is not to replace interpretation.

It is to make subtle findings more visible and longitudinal change easier to recognize.

19. Ethical and Clinical Risks

AI introduces new risks as well as potential benefits.

Automation Bias

A radiologist may accept an AI result without adequate independent review.

Alert Fatigue

Too many false-positive alerts can cause users to ignore the system.

Algorithmic Drift

Performance can change after:

scanner upgrades,
protocol changes,
population changes,
software updates,
or new disease patterns.

Responsibility

Hospitals must establish clear responsibilities for:

AI validation,
human review,
reporting,
follow-up,
incident investigation,
and software updates.

AI governance is therefore inseparable from clinical safety.

20. The Importance of Follow-up Tracking

One of the most underappreciated components of lung nodule management is follow-up.

A hospital can have:

excellent CT scanners,
highly trained radiologists,
sophisticated AI,

and still experience diagnostic delay if follow-up recommendations disappear into the electronic record.

The ideal system creates a closed loop:

Finding → Recommendation → Patient notification → Appointment → Repeat CT → Comparison → Clinical action

This is fundamentally different from simply inserting an AI result into PACS.

21. Five-Year Technology Outlook

Near Term: Automated Triage

AI will increasingly prioritize cases based on:

suspicious nodules,
new lesions,
interval growth,
and potentially other thoracic abnormalities.

Medium Term: Multimodal Risk Modeling

Future systems may integrate:

CT + clinical history + laboratory data + smoking history + genomics

to provide individualized risk estimation.

Longer Term: Continuous Imaging Surveillance

The most significant evolution may not be autonomous diagnosis.

It may be continuous longitudinal monitoring of imaging data.

Instead of asking:

“What is on today's CT?”

the system may ask:

“What has changed in this patient's thoracic imaging over the last five years?”

That is a fundamentally different model of radiology.

22. Ten Expert-Level Insights

1. Busy workflow is a major risk factor for perceptual error.

Even excellent radiologists work within finite attentional resources.

2. Small lesions are inherently difficult.

A 4-mm nodule has very different visual characteristics from a 14-mm mass.

3. Prior imaging is extremely valuable.

Growth can transform a subtle finding into a clinically meaningful one.

4. Detection is not enough.

Tracking may generate greater clinical value than simply finding another candidate nodule.

5. Follow-up is part of diagnosis.

A recommendation that is never completed is not a closed clinical loop.

6. AI should reduce uncertainty, not manufacture certainty.

The algorithm's output should remain subject to expert review.

7. False positives matter.

An AI system that overwhelms radiologists with alerts may create a new workflow problem.

8. Smaller hospitals may have different economics.

Institutions with limited thoracic imaging expertise may derive value from decision support, but the business case must be demonstrated locally.

9. Medico-legal risk is not automatically eliminated by AI.

AI introduces its own governance and documentation requirements.

10. Enterprise integration is more important than the algorithm alone.

A highly accurate algorithm that is disconnected from reporting and follow-up may have limited real-world impact.

23. Clinical Pearls

Pearl 1: A small pulmonary nodule should be interpreted in clinical and longitudinal context.

Pearl 2: Thin-section CT improves characterization of small nodules.

Pearl 3: Always consider comparison with previous CT examinations.

Pearl 4: Interval growth can substantially change clinical interpretation.

Pearl 5: A clear follow-up recommendation is only the beginning of the follow-up process.

Pearl 6: AI should function as an additional safety layer rather than an autonomous diagnostic authority.

24. Common Pitfalls

Pitfall 1: Vessel–Nodule Confusion

A small vessel viewed in cross-section can resemble a pulmonary nodule.

Pitfall 2: Thick-Slice Interpretation

Thicker reconstructions can reduce characterization of small lesions.

Pitfall 3: No Prior Comparison

A new lesion may be incorrectly interpreted as longstanding if previous imaging is not reviewed.

Pitfall 4: AI Overconfidence

A highlighted lesion is not necessarily malignant.

Pitfall 5: Ignoring the Follow-up Chain

The most sophisticated AI cannot compensate for a broken clinical follow-up system.

25. Frequently Asked Questions

Q1. Can AI replace a radiologist?

No.

Current pulmonary AI systems are best understood as decision-support tools that assist detection, measurement, comparison, prioritization, and workflow.

The final interpretation requires clinical context and professional judgment.

Q2. Does AI eliminate missed lung cancer?

No.

AI can potentially reduce certain types of perceptual and workflow-related misses, but it can also generate false positives and false negatives.

Q3. Is FDA clearance proof that an algorithm is the most accurate?

No.

A 510(k) clearance is based on substantial equivalence and applicable safety/performance evidence, not a universal ranking of competing products.

Q4. Can AI save more than $100,000?

A particular case or institutional program may generate savings of that magnitude, but it should not be presented as a universal outcome.

The economic result depends on the number of clinically meaningful delays prevented and the associated downstream costs.

Q5. What is more important: detection or follow-up?

Both are necessary.

The strongest architecture connects detection with longitudinal tracking and closed-loop follow-up.

26. Key Takeaways

The missed pulmonary nodule is not simply a radiologist problem.

It is a clinical systems problem.

The complete pathway is:

Detection → Interpretation → Reporting → Communication → Follow-up → Diagnosis → Treatment

AI can strengthen several points in this chain.

Its most promising functions include:

pulmonary nodule detection,
automated measurement,
prior-study comparison,
growth analysis,
case prioritization,
and follow-up support.

The economic argument should also be framed carefully.

There is no universal rule that:

“AI saves $100,000 per missed cancer.”

Instead, hospitals should calculate:

clinical benefit + avoided downstream cost + workflow efficiency − technology and implementation cost

using their own patient population and financial data.

The real value of medical AI may therefore be much broader than detection.

It may lie in creating a closed-loop imaging intelligence system in which no clinically meaningful finding simply disappears after the radiologist signs the report.

Conclusion

A 4-mm pulmonary nodule may occupy only a tiny fraction of a CT examination.

Yet that tiny finding can become the beginning of a completely different clinical trajectory.

Six months later, a larger and more suspicious lesion may force the healthcare system to ask a difficult question:

Was the cancer always there, and could the system have recognized it earlier?

Modern AI provides one possible answer.

Not by replacing the radiologist.

Not by guaranteeing zero missed cancers.

And not by automatically producing $100,000 of savings.

Its real potential lies in adding an intelligent safety layer across the entire imaging pathway:

CT acquisition → AI detection → radiologist interpretation → longitudinal comparison → structured reporting → follow-up tracking → clinical action.

The future of pulmonary AI is therefore not simply “AI that finds nodules.”

It is AI that helps healthcare systems remember them.

References

[1] F. Bray, M. Laversanne, H. Sung, J. Ferlay, R. L. Siegel, I. Soerjomataram, and A. Jemal, “Global cancer statistics 2022: GLOBOCAN estimates of incidence and mortality worldwide for 36 cancers in 185 countries,” CA: A Cancer Journal for Clinicians, 2024. doi: 10.3322/caac.21834.

[2] National Lung Screening Trial Research Team, “Reduced lung-cancer mortality with low-dose computed tomographic screening,” New England Journal of Medicine, vol. 365, no. 5, pp. 395–409, 2011. doi: 10.1056/NEJMoa1102873.

[3] D. Ardila, A. P. Kiraly, S. Bharadwaj, et al., “End-to-end lung cancer screening with three-dimensional deep learning on low-dose chest computed tomography,” Nature Medicine, vol. 25, pp. 954–961, 2019. doi: 10.1038/s41591-019-0447-x.

[4] A. A. A. Setio, F. Ciompi, G. Litjens, et al., “Pulmonary nodule detection in CT images: False positive reduction using multi-view convolutional networks,” IEEE Transactions on Medical Imaging, vol. 35, no. 5, pp. 1160–1169, 2016. doi: 10.1109/TMI.2016.2536809.

[5] S. G. Armato III, G. McLennan, L. Bidaut, et al., “The Lung Image Database Consortium (LIDC) and Image Database Resource Initiative (IDRI): A completed reference database of lung nodules on CT scans,” Medical Physics, vol. 38, no. 2, pp. 915–931, 2011. doi: 10.1118/1.3528204.

[6] H. MacMahon, D. P. Naidich, J. M. Goo, et al., “Guidelines for management of incidental pulmonary nodules detected on CT images: From the Fleischner Society 2017,” Radiology, vol. 284, no. 1, pp. 228–243, 2017. doi: 10.1148/radiol.2017161659.

[7] A. A. Bankier, H. MacMahon, J. M. Goo, G. D. Rubin, C. M. Schaefer-Prokop, and D. P. Naidich, “Recommendations for measuring pulmonary nodules at CT: A statement from the Fleischner Society,” Radiology, vol. 285, no. 2, 2017. doi: 10.1148/radiol.2017162894.

[8] J. Bueno, L. Landeras, and J. H. Chung, “Updated Fleischner Society guidelines for managing incidental pulmonary nodules: Common questions and challenging scenarios,” RadioGraphics, vol. 38, no. 5, 2018. doi: 10.1148/rg.2018180017.

[9] S. M. Jeon, J.-W. Kwon, S. H. Choi, and H.-Y. Park, “Economic burden of lung cancer: A retrospective cohort study in South Korea, 2002–2015,” PLOS ONE, vol. 14, no. 2, 2019, Art. no. e0212878. doi: 10.1371/journal.pone.0212878.

[10] U.S. Food and Drug Administration, “Medical Device Safety and the 510(k) Clearance Process,” FDA, 2024–2026.

[11] U.S. Food and Drug Administration, “Premarket Notification 510(k),” FDA, 2022–2026.

[12] National Cancer Institute, “Lung and Bronchus Cancer — Cancer Stat Facts,” Surveillance, Epidemiology, and End Results Program, updated 2026.