PET/MRI Radiomics in Patients With Brain Metastases

Philipp Lohmann, Martin Köcher, M Ruge, Veerle Visser‐Vandewalle, N. Jon Shah, Gereon R. Fink, Karl‐Josef Langen, Norbert GalldiksFrontiers in Neurology 2020 · DOI

TL;DR

This review explores the burgeoning field of radiomics, which uses advanced image analysis to extract quantitative features from medical scans (PET and MRI) of brain metastases. By leveraging machine learning, radiomics aims to unlock deeper insights into tumor biology, aiding in diagnosis, prognosis, and treatment response prediction, ultimately enhancing patient care beyond traditional visual interpretation.

What Problem Does This Paper Address?

Brain metastases, secondary tumors in the brain originating from cancers elsewhere in the body, are a common and serious complication. Diagnosing and managing them presents significant challenges. Conventional imaging techniques like MRI, while excellent for visualizing anatomy, often struggle to differentiate between various conditions, such as distinguishing between tumor recurrence and treatment-related changes (like radiation necrosis). This ambiguity can lead to misdiagnosis, potentially resulting in inappropriate or premature termination of effective treatments, with severe consequences for patients. Furthermore, identifying the origin of a brain metastasis when the primary tumor is unknown, and predicting how a patient will respond to different therapies, are critical clinical questions that current imaging methods alone cannot fully answer. Radiomics offers a potential solution by extracting a wealth of quantitative information from medical images that is imperceptible to the human eye, aiming to provide more objective and predictive insights.

Methods

The paper discusses two primary approaches to radiomics:

  • Feature-Based Radiomics: This traditional method involves a multi-step process. First, medical images undergo pre-processing (e.g., normalization, smoothing) to ensure consistency. Then, the region of interest (the tumor) is segmented, often manually, though automated deep learning methods are emerging. Next, a wide array of quantitative features are extracted. These features are broadly categorized into:
  • Shape features: Geometric properties of the tumor.
  • Histogram-based features (first-order statistics): Describing the distribution of pixel intensities within the tumor.
  • Textural features (second-order statistics): Quantifying the heterogeneity and spatial relationships of intensities within the tumor.
  • Higher-order statistics features: Derived from mathematical transformations of the image data. Finally, feature selection is employed to identify the most relevant features, which are then used to train machine learning models (e.g., support vector machines, random forests) for prediction or classification.

  • Deep Learning-Based Radiomics: This approach bypasses the explicit feature extraction and segmentation steps. Instead, artificial neural networks (specifically, deep learning models) are trained directly on the raw or pre-processed image data. These networks automatically learn to identify and extract relevant features at various levels of abstraction, effectively performing feature extraction and model generation simultaneously. This can lead to more robust and potentially more accurate models, especially with large datasets. The paper also briefly mentions transfer learning as a technique to leverage pre-trained neural networks for tasks with limited data.

Datasets

The paper reviews numerous studies that utilized existing clinical datasets. These datasets primarily consist of MRI and PET scans from patients with brain metastases. Specific details on dataset sizes and origins are provided for individual studies within the text and summarized in Table 1. The paper does not introduce any new benchmark datasets.

Key Insights and Ideas

  • Beyond Visual Interpretation: Radiomics fundamentally shifts the paradigm from subjective visual assessment of medical images to objective, quantitative analysis, unlocking a vast amount of information previously hidden within the data.
  • Differentiating the Indistinguishable: Radiomics shows significant promise in distinguishing between treatment-related changes (like radiation necrosis) and actual tumor recurrence, a critical challenge in post-treatment follow-up. This could prevent unnecessary interventions and improve patient management.
  • Predicting the Unknown: The ability of radiomics to predict the origin of brain metastases when the primary tumor is unknown is a significant advancement, potentially guiding further diagnostic workup.
  • Personalized Treatment Insights: Radiomics can offer insights into treatment response, helping to stratify patients and potentially guiding the selection of therapies like immunotherapy or stereotactic radiosurgery.
  • PET/MRI Synergy: Combining data from different imaging modalities, such as PET and MRI, appears to offer superior diagnostic accuracy compared to using either modality alone, suggesting a powerful synergistic effect.
  • Deep Learning's Potential: Deep learning-based radiomics offers a more automated approach, potentially reducing manual labor and improving feature discovery, though it typically requires larger datasets.

Key Results

The review highlights several studies demonstrating high performance metrics:

  • Treatment-Related Changes vs. Recurrence: Studies reported AUCs as high as 0.81 (Peng et al.) and 0.89 (Lohmann et al. combining FET PET/MRI) for differentiating treatment-related changes from recurrence.
  • Predicting Metastasis Origin: Radiomics models achieved AUCs of 0.95 and 0.96 for predicting the origin of brain metastases from specific primary cancers (e.g., lung vs. melanoma).
  • Differentiating Metastases from Glioblastoma: Models demonstrated high accuracy, with AUCs of 0.90 (Qian et al.) and 0.96 (Artzi et al.), outperforming experienced neuroradiologists in some cases.
  • Treatment Response Prediction: Studies reported AUCs of 0.86 (Cha et al. using CT-based deep learning) and identified specific imaging features (e.g., enhancing tumor volume, mean LoG edge feature) associated with improved survival and progression-free survival.

Limitations and Future Work

  • Need for External Validation: A recurring limitation across many studies is the lack of independent external validation datasets. While models may perform well on internal data, their generalizability to different patient populations, scanners, and protocols needs rigorous testing.
  • Standardization Challenges: The field currently lacks standardization in radiomics analysis tools, preprocessing steps, and feature extraction methods. This hinders reproducibility and comparability of results across different research groups.
  • Data Requirements: Deep learning approaches, while promising, often require large, well-curated datasets, which can be challenging to acquire in neuro-oncology due to patient cohort sizes.
  • Interpretability: While radiomics models can be highly predictive, understanding the biological basis of the extracted features and how they relate to specific tumor characteristics remains an active area of research.
  • Future Directions: The authors emphasize the need for standardization, further external validation, and the integration of radiomics into clinical workflows. They also highlight the potential for radiogenomics (linking radiomic features to genetic mutations) and the continued development of automated deep learning techniques.

Summary updated

Abstract

Although a variety of imaging modalities are used or currently being investigated for patients with brain tumors including brain metastases, clinical image interpretation to date uses only a fraction of the underlying complex, high-dimensional digital information from routinely acquired imaging data. The growing availability of high-performance computing allows the extraction of quantitative imaging features from medical images that are usually beyond human perception. Using machine learning techniques and advanced statistical methods, subsets of such imaging features are used to generate mathematical models that represent characteristic signatures related to the underlying tumor biology and might be helpful for the assessment of prognosis or treatment response, or the identification of molecular markers. The identification of appropriate, characteristic image features as well as the generation of predictive or prognostic mathematical models is summarized under the term radiomics. This review summarizes the current status of radiomics in patients with brain metastases.
BibTeX
@inproceedings{lohmann2020pet,
  author = {Philipp Lohmann and Martin Köcher and M Ruge and Veerle Visser‐Vandewalle and N. Jon Shah and Gereon R. Fink and Karl‐Josef Langen and Norbert Galldiks},
  title = {{PET/MRI Radiomics in Patients With Brain Metastases}},
  booktitle = {Frontiers in Neurology},
  year = {2020},
  doi = {10.3389/fneur.2020.00001},
}
Figure