Artificial intelligence (AI) has transformed healthcare by enabling more precise and efficient diagnostics. AI is showing significant promise in interpreting chest CT scans, especially for detecting incidental findings beyond the lungs, such as cardiovascular abnormalities. Chest CT scans, widely used for lung cancer screening, capture more than just pulmonary images, offering a wealth of data on other organs and structures within the thoracic cavity. This incidental information, however, is often overlooked due to the primary focus on lung conditions. A recent study conducted by Marcinkiewicz et al. aims to fill this gap by utilizing AI to identify these incidental findings and predict all-cause mortality (ACM). By integrating AI-driven segmentation of multiple organs and structures from chest CT images, this research sheds light on how this technology can assist radiologists in detecting life-threatening conditions, thereby potentially improving patient outcomes.
The Growing Importance of Extrapulmonary Findings
Thoracic CT scans are commonly performed for various diagnostic purposes, particularly lung cancer screening. However, these scans also capture other vital structures, such as the heart, arteries, and tissues surrounding the lungs, which may hold important diagnostic clues. Extrapulmonary incidental findings, including cardiovascular diseases, are common and can be significant predictors of mortality. As the study highlights, coronary artery calcium (CAC) and abnormal epicardial adipose tissue (EAT) are examples of extrapulmonary findings strongly linked to cardiovascular risks. Unfortunately, these findings are often underreported or neglected, despite their clinical importance.
Marcinkiewicz et al.’s study highlights the potential of AI to not only assist in identifying such incidental findings but also to predict long-term mortality risk. By focusing on 32 segmented structures visible on chest CT scans, the research uses a multistructure AI model that extracts radiomic features from each organ, thus providing a comprehensive evaluation of patient health risks. This goes beyond the traditional approach of manually assessing lung abnormalities, suggesting a paradigm shift in interpreting chest CT scans.
AI Integration in Radiology: Performance and Predictive Power
The study conducted by Marcinkiewicz et al. employed AI models to analyze chest CT scans from a large dataset gathered from the National Lung Screening Trial (NLST). The AI models segmented multiple structures and extracted features that were analysed for their potential to predict two-year and ten-year mortality rates. The study’s findings demonstrate the effectiveness of AI in predicting both short-term and long-term all-cause mortality. The AI model achieved an area under the curve (AUC) of 0.72 for ten-year ACM and 0.71 for two-year ACM, which are promising results that highlight the system's predictive power.
Moreover, CAC was found to be the most important feature for predicting mortality, which aligns with existing clinical knowledge about the role of coronary artery disease in premature death. The study also demonstrated that the AI models could detect significant incidental extrapulmonary findings with high accuracy. These findings suggest that AI can significantly enhance the value of chest CT scans by providing additional, clinically meaningful information that radiologists may overlook due to the overwhelming amount of data.
Incorporating AI-based tools into radiology workflows can improve the accuracy of mortality predictions and assist in identifying other underlying health risks. This advancement has the potential to transform the standard lung cancer screening into a broader diagnostic tool capable of detecting multiple high-risk conditions early on, thus enabling more timely and targeted interventions.
Challenges and Limitations of AI in Chest CT Interpretation
While the study highlights the promise of AI in medical imaging, it also acknowledges certain limitations and challenges. One of the primary issues encountered in the study was variability in reporting incidental findings across different screening sites. The study noted significant discrepancies in how frequently extrapulmonary findings, such as cardiovascular abnormalities, were reported. For example, at some sites, cardiovascular findings were reported in 53% of participants, while at others, no cardiovascular abnormalities were noted. This variability reflects the need for more standardized and automated methods, such as AI-driven analyses, to ensure consistency and reliability in reporting.
Another challenge is the model’s reliance on radiomic features, which, while clinically interpretable, are not as widely recognized as standardized clinical measures. This can pose a barrier to adopting AI models in clinical practice, as radiologists and clinicians may need additional training and confidence in interpreting these AI-generated outputs. Furthermore, due to current limitations in AI segmentation models, the study did not include specific structures, such as lymph nodes and breast tissues. This suggests that while AI technology has advanced significantly, there is still room for improvement in comprehensiveness and accuracy.
The AI model also faced challenges in predicting incidental findings reported by radiologists, as there was a significant drop in prediction performance during external testing. This could be attributed to inconsistencies in how incidental findings are reported, underscoring the need for more standardized reporting frameworks to fully leverage AI's potential in radiology.
Conclusion
The integration of AI in analyzing chest CT scans, as demonstrated by Marcinkiewicz et al., represents a significant advancement in diagnostic imaging. By leveraging AI to segment and analyze multiple structures in the thoracic region, this study shows how incidental findings, often overlooked in traditional radiology, can offer valuable predictive insights into patient mortality. The AI model’s ability to predict both short-term and long-term all-cause mortality based on extrapulmonary findings indicates a future where chest CT scans can serve as a more holistic diagnostic tool. However, challenges remain, including variability in human reporting and the need for more refined AI models capable of handling a broader range of anatomical structures.
As AI technology continues to evolve, its role in healthcare is poised to expand. With further development and validation, AI-driven tools may soon become indispensable in radiology, enabling faster, more accurate diagnoses and potentially improving patient outcomes through earlier intervention. This study marks an important step toward realising AI's full potential in transforming chest CT interpretation, but it also highlights the ongoing need for collaboration between AI developers, radiologists, and clinicians to ensure the successful integration of these tools into everyday practice.
Source Credit: Radiology
Image Credit: iStock