When a patient undergoes a medical scan, the resulting image is typically treated as a targeted answer to a narrow clinical question. A chest CT might be ordered to stage lung cancer; a frontal X-ray to check for pneumonia. Once that primary question is answered, the rest of the image data is effectively discarded.
However, a medical image is not just a picture; it is a high-dimensional physiological profile. In a pilot study, Prediction of future healthcare expenses of patients from chest radiographs using deep learning: a pilot study, the authors demonstrated that standard chest radiographs contain latent indicators of systemic health strong enough to predict future healthcare expenditures.
The failure to leverage this background data creates an opportunistic screening gap. For example, a retrospective study by Maximilian T. Löffler found that while nearly one in three elderly patients undergoing routine CT scans had an osteoporotic vertebral fracture, radiologists explicitly noted these fractures in only 25% of final reports (Epidemiology and reporting of osteoporotic vertebral fractures in patients with long-term hospital records based on routine clinical CT imaging). The data was captured, but the diagnostic insight was lost because it fell outside the primary clinical focus.
Why AI Unlocks Latent Biomarkers
Human vision is inherently task-oriented and optimized for figure-ground separation. A radiologist focusing on a pulmonary nodule will naturally filter out the calcium in the coronary arteries or the density of the surrounding vertebrae. Deep learning models, by contrast, process the entire pixel matrix uniformly. They do not get distracted by the primary indication and can extract multi-system biomarkers simultaneously.
Consider the case of osteoporosis. Traditional screening relies on Dual-energy X-ray absorptiometry (DXA), which requires a dedicated appointment and additional radiation. However, routine CT scans contain rich 3D volumetric data that can be repurposed. Maximilian T. Löffler et al. demonstrated that opportunistic quantitative CT (QCT) derived from routine scans significantly outperforms DXA in predicting future spinal fractures (Improved prediction of incident vertebral fractures using opportunistic QCT compared to DXA).
By automating this process with deep learning, Maximilian T. Löffler et al. built a pipeline that measures 3D bone mineral density without manual intervention, catching high-risk patients who would otherwise be missed (Automatic opportunistic osteoporosis screening in routine CT: improved prediction of patients with prevalent vertebral fractures compared to DXA). Yannik Leonhardt et al. validated this in a prospective cohort, showing that opportunistic measurements outperformed DXA over a three-year follow-up period (Opportunistic QCT Bone Mineral Density Measurements Predicting Osteoporotic Fractures: A Use Case in a Prospective Clinical Cohort). Beyond general fracture prediction, Maximilian T. Löffler et al. demonstrated that this automated bone density data can be used for surgical risk stratification, predicting hardware failure and screw loosening prior to lumbar spinal surgery (Opportunistic Osteoporosis Screening Reveals Low Bone Density in Patients With Screw Loosening After Lumbar Semi-Rigid Instrumentation: A Case-Control Study).
Furthermore, this screening is no longer constrained to traditional anatomical sites like the lumbar spine. Constanze Ramschütz et al. established that opportunistic QCT can reliably assess bone mineral density using the cervical and thoracic spine, expanding the utility of millions of routine scans that do not cover the lower back (Cervicothoracic volumetric bone mineral density assessed by opportunistic QCT may be a reliable marker for osteoporosis in adults).
This paradigm extends far beyond bone health. AI models can look at the exact same routine scans to evaluate entirely different organ systems:
- Cardiovascular Risk: Aryan Zahergivar et al. showed that coronary artery calcium (CAC) scores can be extracted from routine lung cancer staging CTs to predict major adverse cardiovascular events, allowing for integrated cardio-oncology management (Prognostic value of coronary artery calcium scoring in patients with non-small cell lung cancer using initial staging computed tomography).
- Metabolic Conditions: Ayis Pyrros et al. trained a deep learning model to detect Type 2 Diabetes from standard frontal chest radiographs by identifying subtle radiographic markers of adiposity and bone density, outperforming traditional clinical risk scores (Opportunistic detection of type 2 diabetes using deep learning from frontal chest radiographs).
- Oncological Screening: Yingda Xia et al. developed GRAPE, a framework that detects early-stage gastric cancer from routine, noncontrast CT scans originally performed for unrelated reasons, outperforming traditional radiologist assessments (AI-based large-scale screening of gastric cancer from noncontrast CT imaging).
Holistic Phenotyping and Body Composition
Instead of isolating a single disease, opportunistic screening allows for a holistic assessment of a patient's physiological age and tissue composition.
A key driver of this is automated body composition analysis. Routine abdominal CT scans capture detailed cross-sections of muscle, visceral fat, and subcutaneous fat. To unlock this data at scale, Louis Blankemeier et al. introduced Comp2Comp, an open-source Python package that automates the segmentation and quantification of these tissues, overcoming the bottleneck of manual segmentation (Comp2Comp: Open-Source Body Composition Assessment on Computed Tomography).
These extracted metrics provide powerful prognostic signals. For instance, Sung Hye Kong et al. improved fracture prediction models by training a neural network to analyze both vertebral bone structure and the surrounding paravertebral muscle mass simultaneously (A Computed Tomography–Based Fracture Prediction Model With Images of Vertebral Bones and Muscles by Employing Deep Learning: Development and Validation Study). Similarly, Hao Xu et al. utilized routine chest CTs to measure skeletal muscle mass, finding it to be a critical predictor of all-cause mortality in patients with acute exacerbations of chronic obstructive pulmonary disease (The association between skeletal muscle mass and all-cause mortality in acute exacerbation of chronic obstructive pulmonary disease).
Body composition metrics also provide unexpected insights into treatment efficacy. Xin Huang et al. analyzed routine staging scans of patients with platinum-resistant ovarian cancer, discovering that specific adipose tissue areas (visceral and subcutaneous fat) serve as crucial imaging biomarkers that predict a patient's therapeutic response to targeted therapies like apatinib (Adipose tissue area as a predictor for the efficacy of apatinib in platinum-resistant ovarian cancer: an exploratory imaging biomarker analysis of the AEROC trial).
By combining multiple automated biomarkers—such as muscle density and aortic calcification—Perry J. Pickhardt et al. constructed a "biological age" model from routine abdominal CTs that predicts longevity more accurately than chronological age alone (Biological age model using explainable automated CT-based cardiometabolic biomarkers for phenotypic prediction of longevity).
[Routine CT Scan] ───► Primary Diagnosis (e.g., Lung Cancer Staging)
│
▼ (Automated AI Pipeline)
┌───────────────┼───────────────┐
▼ ▼ ▼
[Bone Density] [Muscle Mass] [Vascular Calcium]
│ │ │
└───────────────┼───────────────┘
▼
[Holistic Biological Age] ───► Longevity & Comorbidity Prediction
Eliminating Technical Friction
Historically, extracting quantitative data from routine imaging required specialized protocols or hardware. For example, traditional QCT required placing a physical calibration phantom under the patient during the scan, which is impractical for routine workflows.
To make opportunistic screening viable, recent research focuses on zero-overhead extraction methods. Ferdinand Roski et al. demonstrated that dual-layer spectral CT can accurately measure bone mineral density without any calibration phantoms, leveraging "always-on" spectral data from standard chest or abdomen scans (Bone mineral density measurements derived from dual-layer spectral CT enable opportunistic screening for osteoporosis).
Furthermore, researchers are realizing that data previously dismissed as "technical noise" or "disposable" contains valuable signals. Anna M Marcinkiewicz et al. developed an AI framework that analyzes low-dose, non-gated auxiliary scans—originally used merely for technical attenuation correction in hybrid cardiac imaging—to evaluate extracardiac structures like the lungs and aorta, significantly boosting mortality prediction accuracy (Holistic AI analysis of hybrid cardiac perfusion images for mortality prediction). Similarly, James P. Howard et al. showed that the initial low-resolution "scout" images used for planning cardiac MRIs can be automatically analyzed to detect major abnormalities, allowing the scanner to dynamically optimize subsequent sequences in real-time (Automated analysis and detection of abnormalities in transaxial anatomical cardiovascular magnetic resonance images: a proof of concept study with potential to optimize image acquisition).
Modality-Wide Secondary Utilization
The principle of extracting latent value from routine data applies across all medical imaging modalities, not just CT and X-ray:
- Mammography: Simin Wang et al. demonstrated that high-energy images generated during contrast-enhanced mammography—which are typically unutilized for direct clinical interpretation—contain rich radiomic features that significantly improve breast lesion classification (Radiomic Analysis of Contrast-Enhanced Mammography With Different Image Types: Classification of Breast Lesions).
- Dermatology: Daniel Cho et al. introduced "Dermal Radiomics," applying high-throughput quantitative analysis to skin imaging to extract high-dimensional physiological biomarkers like melanin and hemoglobin that are invisible to the naked eye (Dermal Radiomics for Melanoma Screening).
- Pathology: Tianhua Tan et al. showed that standard Hematoxylin and Eosin (H&E) stained slides can be analyzed at the individual cell level using machine learning to quantify lymphocyte infiltration, creating a novel immune classification system that predicts cancer recurrence (Novel immune classification based on machine learning of pathological images predicts early recurrence of hepatocellular carcinoma).
Economic and Clinical Implications
The core value proposition of opportunistic screening is that it provides "free" diagnostic insights—requiring zero additional radiation exposure, zero extra patient visits, and minimal incremental cost.
Sebastian Rühling et al. evaluated the economic impact of this approach, demonstrating that opportunistic QCT-based osteoporosis screening is highly cost-effective and, in women, actually cost-saving to the healthcare system by preventing debilitating fractures before they occur (Cost-effectiveness of opportunistic QCT-based osteoporosis screening for the prediction of incident vertebral fractures).
By shifting from a reactive diagnostic model to an automated, proactive screening model, healthcare systems can systematically identify multi-organ risks. As outlined by Mikhail Goncharov et al. in the ARILIS study protocol, deploying multi-targeted AI across routine scans allows clinicians to catch hidden cardiovascular, pulmonary, and bone diseases in high-risk populations, such as cancer patients, where non-cancer comorbidities are frequently the leading cause of death (Using artificial intelligence to predict and prevent non-cancer mortality in patients with cancer: ARILIS study protocol).
Key Foundational Work
Dermal Radiomics for Melanoma Screening (2015) — contributed 1 concept elements to this direction.
Simultaneous multiple volume (SMV) acquisition algorithm for real-time navigator gating (2003) — contributed 1 concept elements to this direction.