Longitudinal lung cancer prediction convolutional neural network model improves the classification of indeterminate pulmonary nodules

Rafael Paez, Michael N. Kammer, Aneri Balar, Dhairya A. Lakhani, Michael Knight, Dianna J. Rowe, David Xiao, Brent E. Heideman, Sanja Antic, Heidi Chen, Sheau‐Chiann Chen, Tobias Peikert, Kim L. Sandler, Bennett A. Landman, Stephen A. Deppen, Eric L. Grogan, Fabien MaldonadoScientific 2023 · DOI

Abstract

Deep learning models like the Lung Cancer Prediction Convolutional Neural Network (LCP CNN) have shown promise in stratifying indeterminate pulmonary nodules (IPNs). However, these models typically rely on single-timepoint snapshots, ignoring the longitudinal data often available in clinical practice. This study investigated whether tracking the change in LCP CNN scores over time improves malignancy prediction. Using a prospective-specimen collection, retrospective-blinded-evaluation (PRoBE) design, 80 subjects with IPNs (48 benign, 32 malignant) were analyzed across at least three consecutive CT scans. While initial LCP CNN scores showed no significant difference between groups, longitudinal trends diverged sharply: benign nodules remained stable, while malignant nodules showed a significant increase ($p < 0.001$). A joint model incorporating both score value and slope achieved a time-dependent AUC of 0.857, outperforming snapshot-based models. These findings suggest that longitudinal radiomic modeling can significantly enhance lung cancer risk assessment.

Introduction

Approximately 1.6 million indeterminate pulmonary nodules (IPNs) are detected annually in the United States. While the majority are benign, the clinical challenge lies in accurately identifying the small percentage of malignancies without subjecting patients to unnecessary, invasive biopsies. Current AI tools, such as the FDA-cleared Optellum LCP CNN, provide a malignancy probability score based on a single scan. However, radiologists traditionally rely on temporal changes—such as growth or morphological shifts—to manage nodules. This study bridges the gap between static AI snapshots and longitudinal clinical practice by evaluating the predictive power of AI score trajectories.

Methodology

The study utilized the Optellum LCP CNN model, which generates a malignancy score from 0 to 100. Researchers applied a "PRoBE" design to a cohort of 80 patients from Vanderbilt University Medical Center and the Nashville VA. Inclusion criteria required nodules between 6–30 mm and a minimum of three consecutive CT scans with a slice thickness of ≤ 1.5 mm.

To analyze the temporal data, the team developed a joint model for longitudinal and time-to-event data. They compared four predictive approaches:

  1. Last value: The most recent AI score only.
  2. Value: The full longitudinal history.
  3. Slope: The rate of score change.
  4. Value + Slope: A combined metric of current risk and trajectory.

Results and Key Insights

Divergent Trajectories

The most significant finding was the clear divergence in score trends between benign and malignant nodules. While individual trajectories can be noisy, the aggregate data revealed that benign nodules typically "flatline" or regress, whereas malignant nodules exhibit a consistent upward "climb" in risk scores.

Individual LCP CNN score trajectories for benign (blue) and malignant (red) nodules over time.

Interestingly, the initial LCP CNN scores were not statistically different between the two groups. This suggests that a single baseline scan may be insufficient for definitive risk stratification, whereas the "delta" (change over time) provides a critical diagnostic signal. The difference in trends was highly significant (difference = 0.111, $p < 0.001$).

Predictive Performance

The joint "Value + Slope" model demonstrated superior predictive power compared to snapshot evaluations. When predicting malignancy 12 months into the future (using 24 months of prior data), the longitudinal model achieved a time-dependent AUC of 0.857, compared to 0.757 for the model using only the most recent score.

Modeled trends of sqrt(LCP CNN Score) over absolute months (A) and months relative to diagnosis (B).

Discussion and Clinical Implications

By quantifying morphological changes that may be too subtle for human detection via standard RECIST measurements, longitudinal AI modeling offers a path toward "dynamic risk" assessment. Clinicians could potentially update a patient's probability of cancer with every new scan, allowing for earlier intervention in cases with rising trajectories and reducing the frequency of biopsies for stable, benign nodules.

While this pilot study is limited by a small sample size (n=80) and retrospective design, it provides a strong proof-of-concept for integrating temporal AI data into clinical workflows. Future work will focus on validating these findings in larger, external cohorts and comparing AI-driven score trends directly against traditional volumetric growth measurements.

Summary updated

Abstract

A deep learning model (LCP CNN) for the stratification of indeterminate pulmonary nodules (IPNs) demonstrated better discrimination than commonly used clinical prediction models. However, the LCP CNN score is based on a single timepoint that ignores longitudinal information when prior imaging studies are available. Clinically, IPNs are often followed over time and temporal trends in nodule size or morphology inform management. In this study we investigated whether the change in LCP CNN scores over time was different between benign and malignant nodules. This study used a prospective-specimen collection, retrospective-blinded-evaluation (PRoBE) design. Subjects with incidentally or screening detected IPNs 6-30 mm in diameter with at least 3 consecutive CT scans prior to diagnosis (slice thickness ≤ 1.5 mm) with the same nodule present were included. Disease outcome was adjudicated by biopsy-proven malignancy, biopsy-proven benign disease and absence of growth on at least 2-year imaging follow-up. Lung nodules were analyzed using the Optellum LCP CNN model. Investigators performing image analysis were blinded to all clinical data. The LCP CNN score was determined for 48 benign and 32 malignant nodules. There was no significant difference in the initial LCP CNN score between benign and malignant nodules. Overall, the LCP CNN scores of benign nodules remained relatively stable over time while that of malignant nodules continued to increase over time. The difference in these two trends was statistically significant. We also developed a joint model that incorporates longitudinal LCP CNN scores to predict future probability of cancer. Malignant and benign nodules appear to have distinctive trends in LCP CNN score over time. This suggests that longitudinal modeling may improve radiomic prediction of lung cancer over current models. Additional studies are needed to validate these early findings.
BibTeX
@inproceedings{paez2023longitudinal,
  author = {Rafael Paez and Michael N. Kammer and Aneri Balar and Dhairya A. Lakhani and Michael Knight and Dianna J. Rowe and David Xiao and Brent E. Heideman and Sanja Antic and Heidi Chen and Sheau‐Chiann Chen and Tobias Peikert and Kim L. Sandler and Bennett A. Landman and Stephen A. Deppen and Eric L. Grogan and Fabien Maldonado},
  title = {{Longitudinal lung cancer prediction convolutional neural network model improves the classification of indeterminate pulmonary nodules}},
  booktitle = {Scientific},
  year = {2023},
  doi = {10.1038/s41598-023-33098-y},
}
Figure