Predicting Future Morphological Changes of Lesions from Radiotracer Uptake in 18F-FDG-PET Images

Ulaş Bağcı, Jianhua Yao, Kirsten Miller-Jaster, Xinjian Chen, Daniel J. MolluraPLoS 2013 · DOI

Abstract

This study introduces a novel computational framework for the automated identification of texture and shape features in lesions using (18)F-FDG-PET imaging. By employing a graph-based image segmentation method, the framework predicts future morphological changes with high accuracy. Evaluating 60 scans from 30 patients across various cancer types, the methodology demonstrated a mean Dice similarity coefficient of 85.75% for segmentation. Results indicate that integrating textural features with traditional Standardized Uptake Values (SUV) significantly improves prediction accuracy for longitudinal changes (Spearman correlation = 0.8715), outperforming simple intensity metrics like $SUV_{max}$.

Introduction: Beyond the "Dumb" Metric

In clinical oncology, Positron Emission Tomography (PET) scans are essential for assessing tumor activity. Traditionally, clinicians rely on the Standardized Uptake Value (SUV), a measure of radiotracer intensity. However, SUV is a limited metric; it captures brightness but ignores the complex spatial patterns (texture) and geometric structures (shape) of a lesion. Furthermore, manual segmentation—the process of outlining a tumor—is time-consuming and prone to inter-observer variability.

To address these challenges, researchers have developed a fully automated, quantitative framework designed to characterize lesions and predict how they will morphologically evolve over time.

Methodology: A Three-Step Automated Pipeline

The proposed framework follows a rigorous three-step process to move from raw PET data to predictive insights.

The three-step automated pipeline: (1) Automated IUR detection, (2) Object delineation using ARW, and (3) Feature extraction.

1. Detection

The system first identifies "Interesting Uptake Regions" (IURs). This automated detection phase locates potential lesions without requiring manual intervention, ensuring a consistent starting point for analysis.

2. Segmentation via Automated Random Walk (ARW)

The core of the delineation process is an Automated Random Walk (ARW) method. This approach treats the image as a graph where pixels are nodes. It calculates the probability of a "random walker" starting from a pixel reaching a "lesion seed" versus a "background seed." This adaptive version uses kernel density estimation to mitigate the noise and low resolution inherent in PET imaging.

3. Feature Extraction

Once segmented, the system extracts 68 distinct imaging features. These are categorized into textural, statistical, and morphological groups to provide a comprehensive "fingerprint" of the lesion.

Categories of extracted features, including descriptive statistics, GLCM, Run length, GMRF, and shape features.

Key Insights: Texture as a Predictor

The study's most significant finding is that $SUV_{max}$—the clinical standard—has almost zero correlation with future volume changes. While it indicates current activity, it fails to predict growth or shrinkage.

In contrast, textural features such as Entropy (measuring disorder) and Maximum Probability were far more effective at predicting longitudinal changes. The researchers discovered that the standard deviation of these features within a lesion often provided better predictive power than the average values, suggesting that internal heterogeneity (or "chaos") is a primary driver of tumor progression.

Results and Performance

The framework was validated using a longitudinal dataset of 30 patients with 13 different types of lesions (e.g., lung cancer, lymphoma, sarcomas).

  • Segmentation Accuracy: The ARW algorithm achieved a Mean Dice Similarity Coefficient of 85.75%, showing high fidelity to expert manual tracings.
  • Predictive Accuracy: By combining textural features with SUV measurements, the model achieved a Spearman correlation coefficient of 0.8715 ($p < 2e-16$) for predicting future morphological changes.
  • Feature Significance: 28 out of 68 extracted features showed a strong correlation with $SUV_{max}$, but textural features specifically outperformed $SUV_{max}$ in longitudinal prediction tasks.

Conclusion and Future Directions

This quantitative approach demonstrates that the "internal architecture" of a tumor, captured through texture and shape analysis, is a vital piece of the diagnostic puzzle. While PET resolution remains a constraint, the integration of these automated features significantly enhances our ability to forecast disease progression. Future work aims to incorporate anatomical data from CT or MRI scans (multi-modal fusion) and conduct larger clinical trials to validate these findings for routine clinical use.

Summary updated

Abstract

We introduce a novel computational framework to enable automated identification of texture and shape features of lesions on (18)F-FDG-PET images through a graph-based image segmentation method. The proposed framework predicts future morphological changes of lesions with high accuracy. The presented methodology has several benefits over conventional qualitative and semi-quantitative methods, due to its fully quantitative nature and high accuracy in each step of (i) detection, (ii) segmentation, and (iii) feature extraction. To evaluate our proposed computational framework, thirty patients received 2 (18)F-FDG-PET scans (60 scans total), at two different time points. Metastatic papillary renal cell carcinoma, cerebellar hemongioblastoma, non-small cell lung cancer, neurofibroma, lymphomatoid granulomatosis, lung neoplasm, neuroendocrine tumor, soft tissue thoracic mass, nonnecrotizing granulomatous inflammation, renal cell carcinoma with papillary and cystic features, diffuse large B-cell lymphoma, metastatic alveolar soft part sarcoma, and small cell lung cancer were included in this analysis. The radiotracer accumulation in patients' scans was automatically detected and segmented by the proposed segmentation algorithm. Delineated regions were used to extract shape and textural features, with the proposed adaptive feature extraction framework, as well as standardized uptake values (SUV) of uptake regions, to conduct a broad quantitative analysis. Evaluation of segmentation results indicates that our proposed segmentation algorithm has a mean dice similarity coefficient of 85.75 ± 1.75%. We found that 28 of 68 extracted imaging features were correlated well with SUV(max) (p<0.05), and some of the textural features (such as entropy and maximum probability) were superior in predicting morphological changes of radiotracer uptake regions longitudinally, compared to single intensity feature such as SUV(max). We also found that integrating textural features with SUV measurements significantly improves the prediction accuracy of morphological changes (Spearman correlation coefficient = 0.8715, p<2e-16).
BibTeX
@inproceedings{bagc2013predicting,
  author = {Ulaş Bağcı and Jianhua Yao and Kirsten Miller-Jaster and Xinjian Chen and Daniel J. Mollura},
  title = {{Predicting Future Morphological Changes of Lesions from Radiotracer Uptake in 18F-FDG-PET Images}},
  booktitle = {PLoS},
  year = {2013},
  doi = {10.1371/journal.pone.0057105},
}
Figure