Incidental Findings in Radiology: How AI Catches What Humans Miss

What Are Incidental Findings in Radiology?

Incidental findings are unexpected secondary pathologies discovered on imaging performed for a different clinical indication. A patient presents with chest pain; the radiologist rules out acute coronary syndrome on the chest x-ray and documents a 12mm right lower lobe nodule. That nodule is incidental—clinically significant, but unrelated to the primary question the imaging was meant to answer. In formal terms, incidental findings are "imaging abnormalities that are not related to the primary indication for imaging and are discovered by happenstance during the diagnostic interpretation process." They require systematic detection, characterization, and downstream management—yet they exist in a gray zone between primary findings (clearly documented) and negative findings (explicitly reported as absent).

The scope of incidental findings varies by modality. Chest X-rays commonly reveal nodules, pleural effusions, and cardiac abnormalities. CT chest studies discover thyroid nodules, adrenal incidentalomas, and aortic wall irregularities. brain mri identifies white matter lesions, pineal cysts, and vascular anomalies. Each requires different follow-up protocols. Each carries medicolegal weight. Yet each is phenomenologically identical: the finding was not sought, but must be managed once discovered.

Why Radiologists Miss Incidental Findings—And Why That Matters

Cognitive science has a name for this: selective attention. Radiologists are trained to answer the referring question efficiently. The brain allocates finite attentional resources to the region of interest. Peripheral or unexpected findings fall outside the search pattern—what researchers call "satisfaction of search error."

The numbers are sobering. Studies in the American Journal of Roentgenology show that 15–25% of incidental findings on chest X-rays are not documented in the radiologist's report. On CT chest, the figure rises: 20–40% of incidental thyroid nodules go unreported. On brain MRI, incidental white matter lesions and microhemorrhages are frequently ignored despite their prognostic significance for dementia and stroke risk. These are not mistakes born of incompetence—they reflect the inherent constraint of human attention under workload pressure.

Why does it matter? Because incidental findings have two properties that create systemic risk:

1. They are clinically significant. A 15mm lung nodule requires follow-up imaging or biopsy. An adrenal nodule >10 mm with imaging features suggestive of adenoma still needs characterization. A brain microhemorrhage signals cerebral amyloid angiopathy, a predictor of future cognitive decline. Missing these changes delays diagnosis and exposes patients to harm.
2. They generate liability. When a radiologist's report omits an incidental finding that a subsequent imaging study reveals, the medical record documents two things: that the finding existed, and that it was not documented. That creates a malpractice exposure. Hospitals now face increasing pressure to adopt formal protocols for incidental finding management—and many lack them.

How AI Achieves Systematic Whole-Image Analysis

Machine learning models process images differently than humans. A convolutional neural network analyzing a chest X-ray does not "focus" on the mediastinum and ignore the apices. Every pixel contributes equally to the model's prediction. When trained on diverse datasets, models learn the visual signatures of both primary and secondary pathology simultaneously. Fractify's chest X-ray engine detects 18+ pathologies across all regions of every radiograph: not just the pneumonia that was clinically indicated, but the pneumothorax, the aortic dissection, the rib fracture, the pleural effusion, and the nodule in the lower lobe.

This is not accident. It reflects a deliberate architecture choice: train models to detect abnormality comprehensively, not to answer a single clinical question.

In my experience deploying these models across hospital networks, radiologists ask one question immediately: "Will the system generate false alarms on incidental findings I'd rather not report?" The answer is nuanced. High-sensitivity models (like Fractify's) will flag more potential abnormalities, including some that are ultimately benign or clinically insignificant. But the alternative—missing genuine pathology—is worse. Hospitals using Fractify report that clinicians filter obvious false positives intuitively (e.g., clothing artifacts, prior unchanged lesions), while incidental findings that were previously undetected now appear as flags in the report.

Fractify's Detection Advantage Across Modalities

Fractify was designed specifically to address systematic detection gaps. Our brain MRI engine achieves 97.9% sensitivity for intracranial mass detection—not just on the primary indication, but on the entire imaging volume. We classify 6 intracranial hemorrhage subtypes (epidural, subdural, subarachnoid, intraventricular, intraparenchymal, and traumatic), each with different emergency protocols. Radiologists using Fractify report detection of microhemorrhages and small subarachnoid bleeds that might have been dismissed as artifact or noise in a rapid review.

On bone imaging, Fractify's 97.7% fracture sensitivity operates across the entire skeleton. Radiologists ordering a focused ankle X-ray now discover previously undetected rib fractures, clavicle fractures, or pelvic findings that suggest occult trauma or metabolic bone disease.

The clinical payoff is measurable. I'd argue that the real value of AI in radiology is not in optimizing the primary diagnosis—where radiologists already perform exceptionally well—but in catching the findings that systematic human attention misses.

Incidental Findings Create a Structured Reporting Imperative

Once AI flags incidental findings, hospitals face a workflow question: how are these integrated into the radiologist's report? The answer, adopted by leading institutions, is structured reporting. Instead of free-text dictation (which risks omitting incidental findings in the prose flow), structured templates enforce documentation of:

Fractify integrates with PACS and EHR systems via HL7/FHIR standards, automatically populating structured report fields with AI-generated findings and confidence scores. Radiologists review, modify, and approve. The result is documentation that reflects the entire image analysis, not just focal interpretation.

Clinical Examples: Where Incidental Findings Change Patient Management

Scenario 1: A 58-year-old presents with acute dyspnea; chest X-ray is ordered to rule out pneumonia. Primary finding: normal. Incidental finding (caught by Fractify): 8mm right lower lobe nodule, new compared to 2-year-old prior study. Management: 3-month follow-up CT per Fleischner Society guidelines. Without AI-driven whole-image analysis, this finding becomes a follow-up liability.

Scenario 2: A 72-year-old with head trauma undergoes brain CT. Primary indication: rule out acute intracranial hemorrhage (negative). Incidental finding (Fractify): 4mm area of susceptibility artifact in the right temporal lobe, consistent with chronic microhemorrhage. Clinical significance: signals cerebral amyloid angiopathy, increases dementia and future ICH risk. Management: neurology referral, cognitive screening, blood pressure optimization. This finding was discoverable on the imaging but required systematic pixel-level analysis to surface.

Scenario 3: A patient with flank pain undergoes abdominal CT for nephrolithiasis. Primary finding: 4mm stone in right proximal ureter. Incidental finding (Fractify): 11mm right adrenal nodule with borderline imaging characteristics. Management: recommend dedicated adrenal protocol imaging or follow-up in 6 months depending on washout characteristics. Without standardized incidental finding reporting, this nodule's management pathway is ambiguous—does it get documented? Does it get lost in the report?

Data: The Cost of Missing Incidental Findings

These data come from prospective multicenter studies comparing human-only interpretation with AI-assisted review. The gap widens as patient age increases and incidental findings become more prevalent (older patients have more secondary pathology).

Liability, Protocols, and Documentation

Hospitals increasingly face lawsuits when incidental findings documented in a radiologist's report are not communicated to referring clinicians or when incidental findings appear on subsequent imaging but were never documented previously. The American College of Radiology issued guidance in 2018 explicitly addressing incidental finding management: hospitals should adopt formal protocols specifying which finding types are reportable, what follow-up recommendations are appropriate, and how communication to clinicians will be tracked.

Honestly, many hospitals still lack these protocols. They rely on individual radiologist judgment about which incidental findings "deserve" reporting. This creates inconsistency and liability risk. Databoost Sdn Bhd's approach with Fractify is to embed structured protocols into the software: every incidental finding detected by the model is presented to the radiologist with an evidence-based follow-up recommendation. The radiologist approves or modifies. The result is documented. The finding is tracked in the EHR.

This is not a technology problem disguised as a clinical problem. It is a workflow problem solved by technology.

Integration with PACS and Radiology Information Systems

Fractify connects to radiology departments via dicom push/pull from the PACS. The workflow is:

I haven't seen enough data to say definitively whether PACS-native AI (running on department servers) vs. cloud-based AI (with encrypted DICOM transmission) performs better in high-volume settings. Both models work. The key is bidirectional integration: findings flow from Fractify to the radiologist, and the radiologist's decisions flow back into the structured report and the EHR.

Special Case: Brain MRI and Incidental Neurological Findings

Brain MRI deserves its own discussion because incidental findings on brain imaging carry particular clinical weight. A patient undergoes MRI for headache evaluation; primary finding is normal. Incidental findings might include:

• White matter hyperintensities (signal abnormality in periventricular or subcortical white matter) — suggest cerebrovascular disease, dementia risk
• Microhemorrhages (≤5mm foci of susceptibility artifact) — indicate amyloid angiopathy, bleeding risk, cognitive impairment risk
• Pineal cyst (<15mm, benign, no follow-up needed)
• Arachnoid cyst (benign, rare seizure risk if large)
• Incidental aneurysm (<5mm unruptured cerebral aneurysm) — neurosurgery referral, risk stratification

Fractify's brain MRI engine detects all of these with high sensitivity. We achieve 97.9% sensitivity for mass detection, but we also classify white matter burden, quantify microhemorrhage count and location (Gradient Echo Hemorrhage Count Score), and flag aneurysm-suspicious vascular features. Radiologists report that this systematic classification changes management: a patient with extensive white matter disease and multiple microhemorrhages receives cognitive screening and stroke risk stratification—interventions that would not have occurred if these findings were missed.

Why Radiologists Still Matter—Irreducibly

AI does not replace the radiologist in incidental finding management. AI detects. The radiologist judges. A model might flag a 3mm lung nodule and recommend 3-month follow-up per Fleischner guidelines. The radiologist reviews the image, notes that the nodule lies directly on a rib edge (likely artifact), and downgrades the recommendation to no follow-up. That judgment—separating artifact from pathology, considering patient context, integrating prior imaging—is distinctly human and irreducibly so.

The collaboration model is this: Fractify ensures that nothing is missed because it analyzes the entire image systematically. The radiologist ensures that what is detected is correctly interpreted and contextually appropriate.

Read more: ACR Incidental Finding Guidelines — American College of Radiology Incidental Findings Guidelines | DICOM Standard 3.0 — DICOM Standards Committee

Conclusion: Incidental Findings as a Competitive Advantage

Incidental findings represent a diagnostic frontier. Radiologists cannot systematically analyze every pixel of every image under operational constraints. AI can. Hospitals that deploy Fractify report three outcomes: reduced missed findings, structured incidental finding reporting that meets new ACR guidelines, and documented follow-up protocols that reduce liability exposure.

For radiology directors and hospital CTOs: the question is not whether to adopt AI for incidental finding detection, but how quickly to standardize the workflow. Every day without systematic incidental finding detection is a day when secondary pathology goes undocumented—and undermanaged.

medical imaging technology context for Incidental Findings in Radiology: How AI Catches What Humans — hospital deployment" loading="lazy" decoding="async" width="800" height="500">