Potential Improvements in Lung Nodule Measurement Precision in Computed Tomography Using Artificial Intelligence Software and Its Impact on Clinical Practice - A Simulation Study

Background

Accurate measurement of lung nodules is crucial for the detection and management of lung cancer. Nodule size is the primary basis for risk classification in existing guidelines. However, manual measurements by different doctors may vary greatly. This study explores the potential improvements of artificial intelligence (AI)-assisted software in lung nodule measurement and its impact on clinical management compared to manual measurements.

Lung nodules are common findings in chest computed tomography (CT), with about 95% being benign, but the rest potentially cancerous requiring further action. The size and growth rate of lung nodules are strongly correlated with their malignancy risk, making accurate measurement of nodule size and growth rate key elements in current lung nodule and lung cancer diagnostic and management pathways.

Article Source

This article was written by multiple researchers including Mubarak Patel (MSc), Peter Auguste (PhD), Jason Madan (PhD), Hesam Ghiasvand (PhD), Julia Geppert (PhD), Asra Asgharzadeh (PhD), Emma Helm (MD), Yen-Fu Chen (PhD), Daniel Gallacher (PhD), etc. The authors are affiliated with the Warwick Medical School Applied Health Research Centre, Coventry University Institute of Health and Wellbeing, Bristol Medical School, and University Hospitals Coventry and Warwickshire NHS Trust Radiology Department in the UK. This article was published in June 2024 in the British Journal of Radiology.

Research Design and Process

Research Process

1. Creating Baseline Population: A baseline cohort of lung nodule patients was created based on nodule size distributions reported in the literature.
2. Simulating Measurement Precision and Accuracy: The study simulated the precision and accuracy of nodule size measurements by AI-assisted software compared to unassisted software and AI alone.
3. Simulating Nodule Growth: Nodule growth was simulated over a 4-year time frame, and management strategies were evaluated according to existing clinical guidelines.

Nodule Types

Lung nodules can be broadly classified into solid nodules and sub-solid nodules, categorized based on their density and other characteristics in CT images. This study simulated 1 million solid nodules and 1 million sub-solid nodules separately, and then combined these two types into a sample containing 939,000 solid nodules.

Simulation Results

Simulated Nodule Size Distribution

The final dataset included 1 million highest-risk nodules with baseline diameters between 3 and 30 millimeters, of which 93.9% were solid nodules.

Nodule Growth Results

Solid nodules followed a Gompertz growth curve, while sub-solid nodules followed a linear growth curve. Sub-solid nodules grew slower on average than solid nodules.

Subject Classification and Monitoring

Based on the readings from different readers, nodules were assigned to different management options, including final management, discharge, or CT monitoring.

1. Overall Detection Rate and Specificity of Cancerous Nodules:

   • AI-assisted measurements classified more cancerous nodules as requiring further management (62.5%), compared to 61.4% for manual measurements.
   • AI-assisted measurements slightly improved the discharge rate for benign nodules (95.8%, compared to 95.4% for manual measurements).

2. Monitoring Time for Non-cancerous Nodules:

   • Manual measurements showed significantly shorter average monitoring times compared to AI-assisted measurements, especially in detecting sub-solid nodules. This indicates that although AI-assisted measurements improved sensitivity, they also increased monitoring time for non-cancerous nodules and patient anxiety.

Simulation Assumptions and Parameters

Log-normal distribution was used to simulate true nodule sizes, and parameters for measuring precision and accuracy of four different readers (consensus reader, AI-only, AI-assisted, and manual) were estimated based on existing data.

Specific Algorithms and Data Analysis

The software used in the simulation was RStudio 4.1.0. Standard Log-normal distribution was used for sampling true nodule sizes, and Gompertz growth curves and linear growth models were used to simulate nodule growth.

Research Results

Main Findings

The simulation results in this study indicate that AI-assisted measurements improved nodule precision and accuracy, with the following main advantages: 1. Improved early detection and classification rates for cancerous nodules, shortened detection time for cancerous nodules, and improved clinical handling efficiency. 2. Although the detection rate of cancerous nodules was improved, AI-assisted processing also increased monitoring time for non-cancerous nodules.

Clinical Application Value

The potential improvement in lung nodule measurement using AI-assisted software provides a basis for further data generation and adjustment of clinical guidelines. Overall, AI technology has the potential to standardize nodule measurements, improve accuracy/detection rates, and simplify clinical decision-making processes. However, this is accompanied by additional monitoring time and potential increased patient anxiety, requiring careful balancing of sensitivity and specificity.

Methodological and Algorithmic Innovation

The AI algorithms used in this study demonstrated differences compared to traditional manual measurements, and further revealed the potential clinical applications and benefits of AI-assisted processing through optimization and simulation of model inputs, while also highlighting the risks of changing specificity and sensitivity of AI systems in complex clinical decision-making.

Conclusion

This simulation study is the first to demonstrate the potential impact of AI-assisted improved precision and measurement on patients with cancerous and benign nodules. AI-assisted systems showed higher efficacy in detecting cancerous nodules and improved the precision of nodule diagnosis to some extent. However, this also led to longer monitoring times for benign nodules, a finding that has important implications for the application of AI in lung nodule diagnosis and management.