Combined FET PET/MRI radiomics differentiates radiation injury from recurrent brain metastasis

Philipp Lohmann, Martin Köcher, Garry Ceccon, Elena K. Bauer, Gabriele Stoffels, Shivakumar Viswanathan, Maximilian I. Ruge, Bernd Neumaier, N. Jon Shah, Gereon R. Fink, Karl‐Josef Langen, Norbert GalldiksNeuroImage 2018 · DOI

TL;DR

This study demonstrates that combining radiomic features from MRI and FET PET scans significantly improves the accuracy of differentiating between radiation injury and recurrent brain metastasis in patients who have undergone radiotherapy for brain tumors. This multimodal approach offers a more robust and potentially cost-effective solution for clinical management compared to using either imaging modality alone.

What Problem Does This Paper Address?

Following radiotherapy for brain metastases, patients can develop either a recurrence of the tumor or radiation injury. Both conditions can present with similar symptoms and imaging findings on standard contrast-enhanced MRI (CE-MRI), making it difficult for clinicians to distinguish between them. This diagnostic challenge can lead to inappropriate treatment decisions, such as administering further radiotherapy to already damaged tissue or delaying necessary chemotherapy for recurrent tumors. While advanced PET imaging techniques like dynamic FET PET have shown promise, they are often more expensive and time-consuming. This paper aims to address this diagnostic dilemma by exploring the potential of radiomics, a quantitative image analysis technique, applied to readily available CE-MRI and static FET PET scans, to accurately differentiate these two conditions.

Methods

The researchers employed a radiomics approach, focusing on textural feature analysis. This involves extracting quantitative features that describe the heterogeneity and spatial patterns within medical images, which are often imperceptible to the human eye but can reflect underlying tissue characteristics.

Here's a breakdown of their methods:

  • Image Acquisition: Standard CE-MRI and static FET PET scans were used. Static FET PET refers to images acquired at a single time point after tracer injection, as opposed to dynamic scans that capture tracer uptake over time.
  • Image Pre-processing: Images were resliced to a uniform resolution. For MRI, two types of filtering were applied to enhance image features: a Laplacian-of-Gaussian (LoG) filter and a 3D wavelet transformation (DWT3). These filters are designed to highlight subtle edges and texture variations.
  • Volume of Interest (VOI) Definition: Regions of interest were manually contoured on CE-MRI by an experienced radiation oncologist, blinded to the lesion's etiology. For FET PET, an automated contouring process was used. In cases with multiple lesions, the largest one was selected.
  • Radiomic Feature Extraction: The LIFEx software was used to extract 42 textural features from both CE-MRI (unfiltered and filtered) and static FET PET images. These features capture various aspects of image texture, including statistical indices, histogram features, and features derived from grey-level co-occurrence matrices (GLCM), neighborhood grey-level different matrices (NGLDM), and grey-level run length matrices (GLRLM).
  • Feature Selection and Model Generation: To avoid overfitting, a two-step feature selection process was employed. First, a Mann-Whitney-U test identified features with statistically significant differences between the recurrent metastasis and radiation injury groups. Then, logistic regression models were built using a maximum of five selected features. This limited number of features is crucial for robustness, especially with smaller datasets.
  • Model Validation: The performance of the generated models was assessed using cross-validation techniques (leave-one-out, 5-fold, and 10-fold CV). Diagnostic accuracy, sensitivity, and specificity were calculated for each model.

Datasets

The study utilized data from 52 patients who had previously undergone radiotherapy for brain metastases and presented with new or progressive contrast-enhancing lesions on MRI.

  • Patient Cohort: The cohort consisted of 52 patients (39 women, 13 men) with a median age of 55 years.
  • Diagnosis: The definitive diagnosis of either recurrent brain metastasis or radiation injury was established through histopathology (19 patients) or clinicoradiological follow-up (33 patients).
  • Treatment History: Patients had received various forms of radiotherapy, with stereotactic radiosurgery being the most common (87%).
  • Time Intervals: The median time between MRI and PET acquisition was 15 days. The median time between radiotherapy and the suspicious MRI was 15 months.

The paper does not introduce a new benchmark dataset but rather re-evaluates existing patient data for this specific radiomic analysis.

Key Insights and Ideas

  • Complementary Information: The study highlights that CE-MRI and FET PET capture different aspects of brain tissue pathology. CE-MRI reflects blood-brain barrier disruption, while FET PET indicates amino acid transporter activity. Combining these distinct information sources leads to a more comprehensive understanding of the lesion.
  • Radiomics as a Powerful Tool: Textural feature analysis, a form of radiomics, can extract subtle quantitative information from medical images that is not apparent through visual inspection, proving highly effective in differentiating complex pathologies.
  • Static FET PET Sufficiency: The findings suggest that static FET PET, which is simpler and less time-consuming to acquire than dynamic FET PET, provides sufficient information for radiomic analysis in this context. This increases the clinical feasibility of the approach.
  • Unfiltered vs. Filtered MRI: Both unfiltered and filtered MRI images contributed valuable textural features, indicating that both raw image intensity variations and enhanced edge information are important for discrimination.
  • Shared Textural Patterns: Interestingly, some textural features that differentiated between recurrent metastasis and radiation injury were found to be significant in both MRI and FET PET. This suggests a potential underlying shared mechanism or pattern of alteration in these distinct pathologies.

Key Results

The study achieved the following diagnostic accuracies:

  • CE-MRI alone: 81% accuracy (67% sensitivity, 90% specificity).
  • FET PET alone: 83% accuracy (88% sensitivity, 75% specificity).
  • Combined FET PET/CE-MRI: 89% accuracy (85% sensitivity, 96% specificity).

The combined model demonstrated superior performance across all cross-validation methods, consistently yielding higher accuracies than either modality used in isolation. The highest specificity (96%) was achieved with the combined model, which is particularly important for avoiding unnecessary treatments for radiation injury.

Limitations and Future Work

  • Limited Sample Size: While 52 patients is a reasonable number for a pilot study, the authors acknowledge that no independent validation cohort was available. The models were validated using cross-validation, which is a standard practice but not a substitute for external validation.
  • Retrospective Nature: The study was retrospective, meaning data was collected from past patient records.
  • Generalizability: The authors suggest that further investigation is warranted to confirm these findings in larger, independent cohorts.
  • Understanding Feature Meaning: A key area for future work is to better understand the specific pathophysiological meaning behind the identified radiomic features. While the study shows what features are discriminative, why they are discriminative requires further exploration to translate these mathematical descriptions into clinical understanding.
  • Clinical Implementation: The authors emphasize the need for further investigation into the practical implementation of radiomics into routine clinical practice.

Summary updated

Abstract

Our findings suggest that combined FET PET/CE-MRI radiomics using textural feature analysis offers a great potential to contribute significantly to the management of patients with brain metastases.
BibTeX
@inproceedings{lohmann2018combined,
  author = {Philipp Lohmann and Martin Köcher and Garry Ceccon and Elena K. Bauer and Gabriele Stoffels and Shivakumar Viswanathan and Maximilian I. Ruge and Bernd Neumaier and N. Jon Shah and Gereon R. Fink and Karl‐Josef Langen and Norbert Galldiks},
  title = {{Combined FET PET/MRI radiomics differentiates radiation injury from recurrent brain metastasis}},
  booktitle = {NeuroImage},
  year = {2018},
  doi = {10.1016/j.nicl.2018.08.024},
}
Figure