TLDR: A new study utilized an AI-powered Natural Language Processing (NLP) system to analyze over 115,000 radiology reports, revealing that Incidental Thyroid Findings (ITFs) are common (7.8%) and often trigger a diagnostic cascade of further imaging, biopsies, and surgeries. This cascade frequently leads to the detection of small, low-risk thyroid cancers, underscoring ITFs as a major contributor to thyroid cancer overdiagnosis. The research highlights the critical need for standardized reporting in medical imaging to improve patient care and reduce unnecessary interventions.
Thyroid cancer rates in the United States have significantly increased over the past three decades, tripling from 5.5 to 13.4 per 100,000 people. While this might seem alarming, the mortality rate has remained low, raising concerns about overdiagnosis. This means that many small, slow-growing cancers are being detected that might never have caused harm, leading to unnecessary medical procedures, treatments, and financial burdens.
A major contributor to this overdiagnosis is the widespread use of advanced medical imaging, such as CT scans, for reasons unrelated to the thyroid. These scans often pick up what are called Incidental Thyroid Findings (ITFs) – unexpected thyroid abnormalities. The number of CT scans performed annually has surged, leading to a parallel increase in these incidental findings. A recent analysis estimated that ITFs are present in about 8.3% of CT scans. Each of these findings can potentially trigger a series of further diagnostic tests, including more imaging, biopsies, and surgeries, which may not always benefit the patient.
Despite their frequency and potential impact, ITFs have been poorly understood. Key information about how often they occur, how they are evaluated, their imaging characteristics, and their long-term outcomes has been difficult to gather. This is largely because the crucial data is buried within the unstructured, narrative text of radiology reports in electronic health records. Manually sifting through these reports for large-scale research is simply not practical.
This is where Artificial Intelligence (AI), particularly Natural Language Processing (NLP) techniques, offers a groundbreaking solution. NLP can automate the analysis of unstructured text, unlocking rich clinical details from radiology reports on an unprecedented scale. Previous NLP efforts in thyroid conditions mostly focused on classifying thyroid nodules as benign or malignant, but they didn’t delve into the detailed characteristics of these nodules or extract radiologists’ recommendations, which are vital for understanding the diagnostic cascade.
A recent study, led by Felipe Larios, Mariana Borras-Osorio, Yuqi Wu, and Juan P. Brito, among others, aimed to fill these gaps. They developed and validated a novel, high-performance NLP pipeline to automatically identify ITFs and, more specifically, to characterize Incidental Thyroid Nodules (ITNs) in detail. This included extracting their radiologic features and management recommendations from various imaging modalities. The system was then deployed across a large, multi-site healthcare network to quantify the link between these incidentally detected findings and subsequent diagnostic interventions and thyroid cancer diagnoses. You can read the full research paper here: Artificial Intelligence–Enabled Analysis of Radiology Reports: Epidemiology and Consequences of Incidental Thyroid Findings.
How the AI System Worked
The researchers conducted a retrospective study involving over 115,000 adult patients at Mayo Clinic sites between 2017 and 2023. They used a two-stage, transformer-based NLP pipeline. The first stage classified whether a report contained any ITF. If positive, the second stage, a named entity recognition (NER) module, extracted structured attributes of ITNs, such as location, size, radiologic features (e.g., density, calcifications), and follow-up recommendations. This AI system demonstrated high performance in both identifying ITFs and characterizing their features.
Key Findings: Prevalence and the Diagnostic Cascade
The study found that ITFs were common, present in 7.8% of patients – nearly 1 in 13. The vast majority of these (92.9%) were classified as ITNs. Several factors were independently associated with ITFs, including female sex, older age, and higher BMI. The type of imaging and body region also played a significant role; for instance, neck imaging with nuclear medicine scans had substantially higher odds of detecting an ITF compared to a CT of the chest.
A crucial finding was the inconsistency in how radiologists reported nodule features. Nodule size was documented in less than half of ITN cases (43.8%), and other features like calcifications or density were reported even less frequently (under 15%). When recommendations were provided (in 26.7% of reports), follow-up ultrasound was the most common advice.
The study clearly showed a pronounced diagnostic cascade following an ITF. Patients with ITFs had significantly higher odds of undergoing a thyroid ultrasound (20.2%), thyroid biopsy (6.1% vs 0.1% in those without ITFs), and thyroidectomy (partial or total). While thyroid cancer remained uncommon overall, it occurred far more often in patients with an ITF (1.2% vs 0.02%), representing over 60-fold higher odds of detection. Most of these detected cancers were small, low-risk papillary thyroid carcinomas, reinforcing concerns about overdiagnosis.
Also Read:
- Advancing Breast Cancer Diagnosis with AI: A Multi-View Language Model Approach
- Medical AI’s Role-Playing: Unpacking How LLMs ‘Think’ in Clinical Scenarios
Implications for Healthcare
This research provides large-scale evidence that ITFs are a primary driver of thyroid cancer overdiagnosis. The findings highlight the urgent need for standardized reporting of cross-sectional imaging. The authors suggest creating a structured lexicon for CT and other imaging modalities, similar to the American College of Radiology Thyroid Imaging Reporting and Data System (ACR TI-RADS) for ultrasound. Such standardization could reduce variability in reporting, facilitate international comparisons, and promote more thoughtful, cascade-aware management of incidental findings.
By leveraging AI, this study has unlocked valuable clinical insights from vast amounts of unstructured data, paving the way for better understanding and managing incidental thyroid findings, ultimately aiming to reduce unnecessary diagnostic cascades while still identifying clinically significant diseases.
