The value of nodal information in predicting lung cancer relapse using 4DPET/4DCT

Heyse Li, Nathan Becker, Srinivas Raman, Timothy C. Y. Chan, Jean-Pierre BissonnetteMedical 2015 · DOI

Abstract

The best classifier for predicting local relapse had only tumor features. In contrast, the best classifier for predicting overall relapse included a node feature. Overall, the methods showed that nodes add value in predicting overall relapse but not local relapse.

Introduction

In the management of locally advanced non-small cell lung cancer (NSCLC), predicting whether a patient will suffer a relapse after chemoradiation is critical for personalizing follow-up care. While clinicians extensively use Positron Emission Tomography (PET) and Computed Tomography (CT) to map out tumors, most predictive models focus almost exclusively on the primary tumor itself.

Regional lymph nodes are often involved in the disease, yet their specific imaging characteristics—such as metabolic intensity or "hot spots"—are frequently overlooked as predictive markers. This study investigates whether "nodal information" extracted from these affected lymph nodes adds significant value when forecasting if and where the cancer might return.

Methodology

The researchers utilized a prospective cohort of 25 patients and employed 4DPET/4DCT imaging. The "4D" aspect is crucial as it accounts for respiratory motion, reducing the "blur" caused by breathing and providing a sharper picture of metabolic activity, quantified as Standardized Uptake Values (SUV).

From these images, 37 features were extracted for both the primary tumor (T) and the nodes (N). These included SUV metrics (mean, median, peak, and maximum), volume metrics of metabolic "hot zones," and Intensity Volume Histograms (IVH) describing the distribution of metabolic activity within the volume.

Definitions of the PET and CT features extracted for the predictive models.

To build the predictive models, the authors used Logistic Regression combined with Spearman’s rank correlation and Recursive Feature Elimination (RFE). Performance was evaluated using Matthew’s Correlation Coefficient (MCC), a metric more robust than simple accuracy for determining if a model performs better than a random guess on small datasets.

Dataset and Clinical Outcomes

The study followed 25 patients with locally advanced NSCLC for a median of 26 months. The models were trained to predict two distinct outcomes:

  1. Local Relapse: Recurrence at the original primary tumor site.
  2. Overall Relapse: Recurrence anywhere in the body.
Axial and coronal slices of 4DPET/4DCT showing tumor (magenta) and nodal (cyan) contours used for feature extraction.

Key Results: The Predictive Power of Nodal Information

The findings revealed a striking divergence in the utility of nodal data depending on the type of relapse being predicted.

Local Relapse

For predicting local recurrence, the primary tumor features remained the most relevant. The best model used 8 tumor features but achieved a low MCC of 0.07. Interestingly, adding nodal information actually degraded the model's performance (dropping the MCC to -0.07), suggesting that nodal data acts as "noise" in the context of local failure.

Overall Relapse

In contrast, nodal features proved to be powerful indicators of systemic risk. While a tumor-only model achieved an MCC of 0.10 for overall relapse, the inclusion of nodal features nearly tripled the predictive performance, reaching an MCC of 0.29.

The most significant individual predictor identified was $Vol^N_{0.5max}$—the volume of the node where metabolic activity is at least 50% of the maximum intensity. This single nodal feature provided a more reliable prediction for overall survival and relapse than complex combinations of tumor-only data.

Discussion and Limitations

The study demonstrates that while the primary tumor provides information about the "local battle," regional lymph nodes act as a "canary in the coal mine" for systemic failure. The metabolic intensity within these nodes captures the aggressive potential of the disease more effectively than tumor size alone.

The primary limitation is the small sample size (n=25), requiring larger multi-center studies for validation. Additionally, the diagnosis of relapse was based on imaging alone rather than pathological biopsies, which can occasionally be misled by post-treatment inflammation. Future work will explore "second-order" radiomic features like texture and entropy, as well as "delta-radiomics" to track how these nodal features change during the course of treatment.

Summary updated

Abstract

The best classifier for predicting local relapse had only tumor features. In contrast, the best classifier for predicting overall relapse included a node feature. Overall, the methods showed that nodes add value in predicting overall relapse but not local relapse.
BibTeX
@inproceedings{li2015value,
  author = {Heyse Li and Nathan Becker and Srinivas Raman and Timothy C. Y. Chan and Jean-Pierre Bissonnette},
  title = {{The value of nodal information in predicting lung cancer relapse using 4DPET/4DCT}},
  booktitle = {Medical},
  year = {2015},
  doi = {10.1118/1.4926755},
}
Figure