Development and Prospective Validation of an Ultrasound Prediction Model for the Differential Diagnosis of Benign and Malignant Subpleural Pulmonary Lesions: A Large Ambispective Cohort Study

Ke Bi, Demeng Xia, Lin Fan, Xiaofei Ye, Yi Zhang, Mengjun Shen, Hongwei Chen, Yang Cong, Hui-ming Zhu, Chunhong Tang, Jing Yuan, Yin WangFrontiers in Oncology 2021 · DOI

TL;DR

This study introduces a novel ultrasound-based prediction model that integrates B-mode and contrast-enhanced ultrasound (CEUS) parameters to accurately differentiate between benign and malignant subpleural pulmonary lesions (SPLs). The model demonstrates superior diagnostic performance compared to existing ultrasound criteria, offering a non-invasive, radiation-free tool for clinical decision-making.

What Problem Does This Paper Address?

Accurately distinguishing between benign and malignant pulmonary lesions, particularly those located just beneath the pleura (subpleural pulmonary lesions or SPLs), is crucial for patient management. While CT is the standard for initial diagnosis, it can sometimes miss atypical lesions, necessitating invasive procedures or prolonged follow-up, which carry risks and increase healthcare costs. Ultrasound (US) offers a real-time, radiation-free alternative. Both B-mode US and CEUS have shown promise in this area, with specific parameters like enhancement patterns and arrival times (AT) being investigated. However, previous studies have often relied on single indicators or small sample sizes, leading to limitations in their diagnostic accuracy and generalizability. This paper aims to address this gap by developing and validating a comprehensive US prediction model that leverages multiple parameters from both B-mode and CEUS to improve diagnostic precision for SPLs.

Methods

The researchers developed a prediction model using a retrospective dataset (development cohort, DC) and then prospectively validated it on a separate group (validation cohort, VC). They collected 18 parameters from B-mode US and CEUS. A binary logistic regression model was built using the DC to identify key predictors. The model's discrimination and calibration were then assessed internally (in the DC) and externally (in the VC). Crucially, the diagnostic performance of the new model was compared against established ultrasound diagnostic criteria, specifically "lesion-lung AT difference ≥ 2.5 s" and "AT ≥ 10 s."

Datasets

The study utilized two datasets:

  • Development Cohort (DC): A retrospective dataset of 592 patients collected from July 2017 to December 2018.
  • Validation Cohort (VC): A prospective dataset of 220 patients collected from January to April 2019.

These datasets were collected from a leading pulmonary hospital, aiming to provide a robust basis for model development and validation.

Key Insights and Ideas

  • Multimodal US is Key: The study highlights that combining information from both B-mode and contrast-enhanced ultrasound provides a more comprehensive picture than relying on single parameters.
  • Time-Related CEUS Parameters are Powerful: Quantitative CEUS parameters, particularly those related to contrast agent arrival times, play a significant role in differentiating lesion types.
  • A Novel "Ratio of AT Difference" Parameter: The introduction of a "ratio of AT difference" parameter, which normalizes arrival times against both lung and thoracic wall, proved to be a valuable addition, potentially overcoming limitations of individual AT measurements.
  • Morphological Clues Matter: Even a simple B-mode parameter like the angle between the lesion border and the thoracic wall was found to be a significant predictor, underscoring the importance of basic morphological assessment.
  • Simplified Classification: The model uses a simplified classification for vascular signs (tree-like vs. curly hair-like) and non-enhancing regions, aiming for better reproducibility compared to more complex systems.

Key Results

The developed prediction model, incorporating 6 key parameters (angle between lesion border and thoracic wall, basic intensity, lesion-lung AT difference, ratio of AT difference, vascular sign, and non-enhancing region type), demonstrated excellent diagnostic performance:

  • High Discrimination: The model achieved a C-statistic of 0.974 in the development cohort and 0.980 in the validation cohort, significantly outperforming existing criteria.
  • Superiority over Existing Criteria: The model's C-statistics were substantially higher than those for "lesion-lung AT difference ≥ 2.5 s" (0.842 in DC, 0.777 in VC) and "AT ≥ 10 s" (0.688 in DC, 0.641 in VC).
  • Improved Sensitivity and Specificity: The model's sensitivity and specificity were higher than the compared criteria in both cohorts. For instance, in the validation cohort, the model achieved 92.86% sensitivity and 92.59% specificity, compared to 80.36% sensitivity and 75.00% specificity for the lesion-lung AT difference criterion.
  • Good Calibration: The calibration curves showed good agreement between predicted and actual malignancy probabilities, indicating reliable risk stratification.

Limitations and Future Work

The authors acknowledge a few limitations:

  • Single-Center Study: The study was conducted at a single, leading pulmonary hospital, which might introduce bias due to a non-representative distribution of disease types. However, they argue that patients from across the country were treated at their center, mitigating this to some extent.
  • Limited B-US Parameters: Not all B-mode US features and color Doppler flow imaging parameters were included, as gas artifacts and respiratory movement can affect their stability.

For future work, the researchers plan to conduct a prospective, multicenter study with an expanded set of US parameters to further refine and improve their model. They believe their model, being simple, non-radioactive, and clinically feasible, has significant potential for widespread use.

Summary updated

Abstract

www.chictr.org.cn, identifier ChiCTR1800019828.
BibTeX
@inproceedings{bi2021development,
  author = {Ke Bi and Demeng Xia and Lin Fan and Xiaofei Ye and Yi Zhang and Mengjun Shen and Hongwei Chen and Yang Cong and Hui-ming Zhu and Chunhong Tang and Jing Yuan and Yin Wang},
  title = {{Development and Prospective Validation of an Ultrasound Prediction Model for the Differential Diagnosis of Benign and Malignant Subpleural Pulmonary Lesions: A Large Ambispective Cohort Study}},
  booktitle = {Frontiers in Oncology},
  year = {2021},
  doi = {10.3389/fonc.2021.656060},
}
Figure