Radiomics and Its Feature Selection: A Review

Wenchao Zhang, Yu Guo, Qiyu JinSymmetry 2023 · DOI

TL;DR

This review paper provides a comprehensive overview of radiomics, a quantitative approach to extracting features from medical images for disease diagnosis and prognosis. It specifically delves into the critical step of feature selection, highlighting its importance in enhancing model performance and generalizability, and discusses various methods and challenges in this rapidly evolving field.

What Problem Does This Paper Address?

Medical imaging plays a crucial role in diagnosing and monitoring diseases, but manual image interpretation can be subjective and time-consuming. Radiomics aims to overcome this by extracting a vast number of quantitative features from medical images, offering objective and reproducible insights into disease characteristics. However, the sheer volume of extracted features can lead to issues like overfitting and reduced model efficiency. This paper addresses the challenge of identifying the most relevant and informative features from this large set, a process known as feature selection, which is vital for building robust and clinically useful radiomics models.

Methods

The paper outlines the standard radiomics workflow, which includes image acquisition and pre-processing, image segmentation, feature extraction, model creation and evaluation, and clinical application. It categorizes radiomics features into two main types: hand-crafted features (derived from traditional algorithms like intensity, shape, and texture) and deep learning (DL) features (extracted by DL models, particularly Convolutional Neural Networks). The review then focuses on feature selection methods, categorizing them into three main types:

  • Filter Methods: These methods evaluate features based on their intrinsic characteristics, independent of any learning algorithm. They are computationally efficient and less prone to bias towards a specific model. Examples include variance, correlation coefficients, and mutual information.
  • Wrapper Methods: These methods use a specific learning algorithm to evaluate subsets of features. They are more computationally intensive but can achieve better predictive accuracy by considering feature interactions. Examples include Recursive Feature Elimination (RFE).
  • Embedded Methods: These methods integrate feature selection directly into the model construction process, offering a balance between filter and wrapper methods. Examples include LASSO and Gradient Boosted Decision Trees (GBDT).

The paper also discusses various search strategies for feature subset generation, including complete search, heuristic search, and random search.

Datasets

The paper does not describe specific datasets used for training or evaluation in its own research. Instead, it reviews existing literature and discusses the application of radiomics and feature selection across various cancer types and other diseases, citing numerous studies that have utilized different imaging modalities (CT, MRI, PET) and patient cohorts.

Key Insights and Ideas

  • Radiomics as a Bridge: Radiomics acts as a crucial bridge between raw medical images and actionable clinical insights, transforming qualitative visual information into quantitative data for predictive modeling.
  • Feature Selection is Paramount: The paper strongly emphasizes that feature selection is not just a preprocessing step but a critical component for radiomics success, directly impacting model performance, generalizability, and clinical utility.
  • The "Black Box" Challenge of DL: While deep learning offers powerful feature extraction capabilities, its "black box" nature can hinder interpretability, posing a challenge for clinical adoption.
  • Data Quality and Standardization are Key: The review highlights the significant impact of data acquisition variability and the need for standardization (e.g., through initiatives like IBSI) to ensure reproducibility and external validity of radiomics models.
  • Radiomics' Broad Applicability: Beyond oncology, radiomics shows promise in diagnosing and monitoring non-tumor conditions, demonstrating its adaptability across diverse medical scenarios.

Key Results

The paper does not present novel experimental results. Instead, it synthesizes findings from a broad literature review. Key qualitative findings include:

  • Radiomics has demonstrated potential in improving or surpassing existing tumor diagnosis methods for various cancers, aiding in the prediction of prognosis, tumor subtypes, and metastasis.
  • Feature selection methods, when applied effectively, can reduce data complexity, enhance model performance, and mitigate overfitting.
  • The choice of feature selection method can significantly influence the outcome, and there is no single "best" method for all scenarios.

Limitations and Future Work

The authors acknowledge several limitations and suggest future directions:

  • Reproducibility and Scalability: Challenges remain in ensuring the reproducibility and scalability of radiomics models, particularly due to data acquisition variability and the need for standardization.
  • Interpretability: The "black box" nature of DL models in radiomics poses a challenge for clinical interpretability, and future work should focus on merging DL strengths with the interpretability of hand-crafted methods.
  • Validation: Many existing radiomics studies are limited by small sample sizes from single institutions. Future research should prioritize validation and refinement through large-scale, randomized controlled clinical trials.
  • Optimal Feature Selection: While various feature selection strategies are discussed, determining the optimal selection for diverse disease types and clinical contexts requires further empirical study.
  • Emerging Trends: The paper points to the potential of analyzing raw imaging data directly, bypassing traditional reconstruction steps, as a novel direction for AI-based diagnostics.

Summary updated

BibTeX
@article{zhang2023radiomics,
  author = {Wenchao Zhang and Yu Guo and Qiyu Jin},
  title = {{Radiomics and Its Feature Selection: A Review}},
  journal = {Symmetry},
  year = {2023},
  doi = {10.3390/sym15101834},
}
Figure