Physicians

What is Radiomics?

Radiomics, situated at the intersection of medical imaging and advanced analytics, entails the extraction and analysis of quantitative features from medical images. This field has seen a transformative synergy with artificial intelligence (AI), especially machine learning and deep learning algorithms. The incorporation of AI into radiomics workflows holds immense promise, revolutionizing medical imaging by boosting precision, efficiency, and diagnostic capabilities.

Clinical applications and case studies

Radiomics has emerged as a transformative approach in medical imaging, offering a wealth of quantitative information extracted from images. The practical applications of radiomics span various clinical scenarios, with notable advancements in oncology, neurology, and other fields. Here, we delve into the practical implications through case studies and research findings.

  • Oncology
  • Tumor Characterization and Subtyping[1]

Radiomics plays a pivotal role in characterizing tumors and identifying subtypes based on imaging features. Case studies in breast cancer have demonstrated the ability of radiomic features to differentiate between molecular subtypes, providing valuable insights for treatment planning.

  • Treatment Response Prediction[2]

In oncology, radiomics aids in predicting treatment responses. For instance, research in lung cancer has shown that radiomic analysis of CT scans can predict responses to chemotherapy and immunotherapy. This enables clinicians to tailor treatment strategies for better outcomes.

  • Prognostic Stratification[3]

Radiomics facilitates prognostic stratification by extracting features associated with disease aggressiveness. In prostate cancer, studies have revealed the potential of radiomics in predicting biochemical recurrence, aiding in the identification of patients who may benefit from more aggressive interventions.

2.2. Neurology

  • Neurodegenerative Disease Diagnosis[4]

Radiomics contributes to the diagnosis of neurodegenerative diseases. Case studies in Alzheimer’s disease have demonstrated the ability to extract subtle changes in brain structure and texture, aiding in early and accurate diagnosis.

  • Stroke Assessment[5]

In neurology, radiomics is employed in stroke assessment. Research findings indicate that radiomic features extracted from MRI and CT scans can assist in evaluating lesion characteristics, guiding treatment decisions, and predicting patient outcomes.

  • Tumor Detection in Brain Imaging[6]

Radiomics has proven valuable in the detection and characterization of brain tumors. Case studies highlight its efficacy in distinguishing between tumor types, grading malignancy, and aiding in surgical planning for improved precision.

  • Beyond Oncology and Neurology

Radiomics has applications in cardiology, where it contributes to the assessment of cardiac function and tissue characteristics. Research findings indicate its potential in identifying subtle myocardial changes and predicting cardiovascular events.

  • Musculoskeletal Imaging[8]

In musculoskeletal imaging, radiomics aids in the assessment of joint diseases and bone tumors. Case studies in arthritis demonstrate the utility of radiomic features in quantifying disease severity and monitoring treatment responses.

  • Pulmonary Diseases[9]
    Radiomics extends its reach to pulmonary diseases, where it assists in the diagnosis and monitoring of conditions such as chronic obstructive pulmonary disease (COPD) and pulmonary fibrosis. Research findings emphasize its role in predicting disease progression and guiding therapeutic interventions

 

[1] Marino M. et al.: Radiomics for Tumor Characterization in Breast Cancer Patients: A Feasibility Study Comparing Contrast-Enhanced Mammography and Magnetic Resonance Imaging https://www.ncbi.nlm.nih.gov/pmc/articles/PMC7400681/

[2] Chang R. et al.: Predicting chemotherapy response in non-small-cell lung cancer via computed tomography radiomic features: Peritumoral, intratumoral, or combined? https://www.ncbi.nlm.nih.gov/pmc/articles/PMC9393703/

[3] Bao J. et al: Prediction of clinically significant prostate cancer using radiomics models in real-world clinical practice: a retrospective multicenter study https://insightsimaging.springeropen.com/articles/10.1186/s13244-024-01631-w/

[4] Bevilacqua R. et al.: Radiomics and Artificial Intelligence for the Diagnosis and Monitoring of Alzheimer’s Disease: A Systematic Review of Studies in the Field https://www.ncbi.nlm.nih.gov/pmc/articles/PMC10455452/

[5] Dragos H. et al.: MRI Radiomics and Predictive Models in Assessing Ischemic Stroke Outcome—A Systematic Review https://www.mdpi.com/2075-4418/13/5/857

[6] Zhou M. et al.: Radiomics in Brain Tumor: Image Assessment, Quantitative Feature Descriptors, and Machine-Learning Approaches https://www.ncbi.nlm.nih.gov/pmc/articles/PMC5812810/

[7] Polidori T. et al.: Radiomics applications in cardiac imaging: a comprehensive review https://www.ncbi.nlm.nih.gov/pmc/articles/PMC10338589/

[8] Fritz B. et al.:Radiomics and Deep Learning for Disease Detection in Musculoskeletal Radiology: An Overview of Novel MRI- and CT-Based Approaches https://pubmed.ncbi.nlm.nih.gov/36070548/

[9] Frix A. et al.: Radiomics in Lung Diseases Imaging: State-of-the-Art for Clinicians https://www.ncbi.nlm.nih.gov/pmc/articles/PMC8306026/

Understanding radiomics workflow

  1. Image acquisition:
    In the realm of radiomics, standardized imaging modalities such as Computed Tomography (CT), Magnetic Resonance Imaging (MRI), and Positron Emission Tomography (PET) scans serve as the cornerstone. These modalities offer diverse perspectives, enabling comprehensive evaluation of tumor morphology and physiology.
  2. Tumor annotation:
    Tumor annotation, a fundamental step in radiomics, encompasses both manual delineation by skilled radiologists and automated segmentation via cutting-edge algorithms. While manual segmentation ensures precision, automated techniques streamline the process, enhancing efficiency and reproducibility.
  3. Feature extraction:

Feature extraction constitutes the essence of radiomics, involving the derivation of quantitative metrics from imaging datasets. These encompass a spectrum of characteristics, including intensity, texture, and shape, elucidating the underlying heterogeneity and complexity of tumors.

  1. Data analysis:

Harnessing the power of machine learning algorithms, radiomics embarks on a transformative journey of data analysis. These algorithms meticulously scrutinize the extracted features, deciphering intricate patterns and relationships to furnish clinically relevant insights imperative for informed decision-making.

In essence, radiomics goes beyond traditional imaging methods to uncover the hidden details of how tumors develop and respond to treatment. By seamlessly fitting into everyday medical practice, radiomics is paving the way for a new era in cancer care, where each patient gets personalized treatment.

Using clinical Data to enhance prediction results

  1. Incorporating clinical data alongside medical imaging data can significantly improve the accuracy and relevance of radiomic models. By combining clinical information with radiomic features, we gain a more comprehensive understanding of disease characteristics and patient outcomes.
    Clinical data includes various patient-specific factors such as:
  • Demographic details like age and gender
  • Medical history, including past medical conditions and comorbidities
  • Results from laboratory tests, such as blood tests and biomarker levels
  • Pathological results like tumor grade and histology
  • Details about treatments received, such as chemotherapy, immunotherapy, or radiation therapy
  • Clinical symptoms and signs, such as pain scores and performance status
  • Genetic or molecular biomarkers, such as gene mutations or protein expression levels
  • Follow-up information, including disease progression and survival outcomes
  • Annotations or interpretations provided by radiologists, such as reports and lesion characteristics.

Advanced feature extraction techniques

  1. Hand-crafted features:

Traditional feature extraction methods, often referred to as handcrafted features, are based on predefined algorithms. These methods involve manually defining and calculating features from the data. Some common examples include:

  • Statistical Features: These are derived from statistical properties of the data, such as mean, median, variance, skewness, and kurtosis.
  • Geometrical Features: These include shape descriptors, Fourier descriptors, and other geometrical properties.
  • Texture Features: These are calculated from the spatial distribution of intensity values in an image
  • Spectral Features: These are derived from the frequency domain representation of the data

 

2. Deep Learning:

Deep learning techniques have revolutionized the field of feature extraction, especially in imaging data. Unlike traditional methods, deep learning automatically learns the most relevant features from the data during the training process. This is achieved through the use of neural networks with multiple layers (hence “deep”). Some common deep learning techniques for feature extraction include:

  • Convolutional Neural Networks (CNNs)
  • Transfer Learning

 

3. Machine Learning and Deep Learning in Radiomics
The fusion of Machine Learning (ML) and Deep Learning (DL) with radiomics has completely transformed how we examine and understand medical images. These cutting-edge technologies provide advanced tools for extracting, analyzing, and interpreting extensive datasets, leading to more precise and predictive findings in healthcare.

  • Supervised Learning
  • Unsupervised Learning
  • Transfer Learning
  • Federated Learning

 

4. Machine learning models
Machine learning models are employed in radiomics for prediction tasks. Here are some commonly used models:

  1. Random Forests
  2. Support Vector Machines (SVM)
  3. Artificial Neural Networks (ANN)

Model validation

Validating radiomics models is crucial in their development to ensure they perform well on new data and can be applied beyond the specific datasets they were trained on. This involves assessing how accurately and reliably the models predict outcomes using different statistical and machine learning methods. Through validation, we can detect potential problems like overfitting, underfitting, and biases in the model, resulting in the creation of more dependable and clinically useful tools.

In radiomics, where models analyze complex medical imaging data to offer insights into diagnosis, prognosis, and treatment response, validation is essential for:

  • Ensuring Reliability: Validating radiomics models against external datasets or using cross-validation techniques ensures that the model’s predictions are reliable and applicable to a broad patient population.
  • Clinical Acceptance: Rigorous validation processes are necessary for clinical acceptance and regulatory approval, as they demonstrate the model’s effectiveness and safety in real-world settings.

Techniques for Model Validation

  1. Holdout Method: The dataset is split into training and testing sets, where the model is trained on the former and validated on the latter. This method is straightforward but may not fully utilize the available data, especially in smaller datasets.[1]
  2. Cross-Validation: The dataset is divided into k smaller sets (or “folds”), and the model is trained and tested k times, each time using a different fold as the test set and the remaining as the training set. Cross-validation, especially k-fold cross-validation, is widely used for its better utilization of data and more reliable estimation of model performance.[2]
  3. Bootstrap Methods: These involve randomly sampling with replacement from the dataset to create many training sets, then training and testing the model on these sets. Bootstrap methods are useful for estimating the variability of model performance.[3]
  4. External Validation: The model is validated on a completely independent dataset, not used during the training process. External validation is the gold standard for assessing the generalizability of radiomics models across different populations and imaging protocols.[4]

Metrics for Evaluation [5]

  • Accuracy, Sensitivity, and Specificity: Common metrics that provide a general sense of model performance in classification tasks.
  • Area Under the ROC Curve (AUC): A comprehensive metric that evaluates the model’s ability to distinguish between classes at various threshold settings.
  • Mean Absolute Error (MAE) and Root Mean Squared Error (RMSE): Metrics used for regression tasks, measuring the average magnitude of errors in the model’s predictions.

Challenges and Solutions

  • Data Diversity and Size: The heterogeneity and limited size of medical imaging datasets can challenge model validation. Solutions include augmenting datasets, using federated learning approaches, and developing models robust to data variations.
  • Model Complexity: The complexity of deep learning models used in radiomics can make validation challenging due to overfitting. Techniques such as regularization, dropout, and early stopping are employed to mitigate this.

[1] https://www.comet.com/site/blog/understanding-hold-out-methods-for-training-machine-learning-models/

[2] https://www.analyticsvidhya.com/blog/2021/05/4-ways-to-evaluate-your-machine-learning-model-cross-validation-techniques-with-python-code/

[3] https://docs.rapidminer.com/latest/studio/operators/validation/bootstrapping_validation.html

[4] https://doi.org/10.1093/ckj/sfaa188

[5] https://www.analyticsvidhya.com/blog/2019/08/11-important-model-evaluation-error-metrics/

Interpretability

In the context of AI models for radiomics, interpretability refers to the ability to comprehend and articulate how these models make decisions. Considering the complexity and high stakes of medical diagnostics and treatment planning, interpretability of AI models is not just a technical requirement but a clinical necessity. It serves as a bridge between sophisticated machine learning algorithms and their practical, transparent application in healthcare settings.

Importance in Radiomics

  • Clinical Trust and Adoption: Interpretability is crucial for building trust among healthcare professionals. When clinicians understand how a model arrives at its predictions or classifications, they are more likely to trust and use it as part of their diagnostic and treatment decision-making process.
  • Model Validation and Improvement: Interpretable models allow developers and clinicians to validate the reasoning behind model predictions, ensuring that these are based on clinically relevant features rather than spurious correlations or biases present in the training data.
  • Regulatory Compliance: Many regulatory bodies now require a degree of transparency and interpretability in AI models used in healthcare to ensure patient safety and model reliability.

Strategies for Enhancing Interpretability

  1. Feature Visualization and Importance: Techniques like saliency maps or feature importance rankings can highlight which parts of an image or which features were most influential in the model’s decision, providing insights into the model’s focus and rationale.
  2. Model Simplification: While simpler models like decision trees inherently offer more interpretability, they often come at the cost of reduced accuracy. Balancing complexity and interpretability is key, and sometimes, simpler models can be sufficient for specific applications in radiomics.
  3. Post-Hoc Interpretation: Techniques applied after model training, such as LIME (Local Interpretable Model-agnostic Explanations)[1] and SHAP (SHapley Additive exPlanations)[2], can offer insights into individual predictions, helping to understand how different features contribute to the model’s output.
  4. Interactive Tools: Developing interactive tools that allow users to modify input data and observe changes in model predictions can help clinicians understand how different radiomic features influence the model’s decision-making process.

Challenges in Interpretability

  • Complexity vs. Interpretability Trade-off: The most accurate models in radiomics, such as deep neural networks, are often the least interpretable due to their complexity. Finding a balance or developing methods to interpret these models remains a significant challenge.
  • Quantifying Interpretability: Unlike performance metrics such as accuracy or precision, interpretability is subjective and harder to quantify, making it challenging to compare the interpretability of different models or to measure improvements in interpretability.

As AI progresses in radiomics, it’s crucial for researchers, clinicians, and data scientists to grasp these basic concepts. Continuously exploring and refining these principles will unlock AI’s potential in transforming medical imaging and personalized healthcare.

[1] https://homes.cs.washington.edu/~marcotcr/blog/lime/

[2] https://datascientest.com/en/shap-what-is-it

Challenges and limitations

Technical challenges:

Radiomics, despite its potential, faces several technical challenges that need to be addressed to ensure its effective implementation.

  • Image Quality: The quality of the images used in radiomics is crucial as it directly impacts the accuracy of the analysis. Poor image quality can lead to inaccurate results, making it essential to use high-quality images and employ techniques to enhance image quality where necessary.
  • Reproducibility: Reproducibility is a significant challenge in radiomics. Different imaging protocols and scanners can lead to variations in the extracted features, affecting the reproducibility of the results. Standardization of imaging protocols and use of robust features can help address this issue.
  • Standardization: There is a lack of standardization in radiomics, particularly in terms of feature definition and extraction methods. This lack of standardization can lead to inconsistent results across different studies. Efforts are being made to develop standardized protocols and guidelines for radiomics analysis.

 

  • Clinical limitations

In addition to technical challenges, there are also several clinical limitations that need to be considered when integrating radiomics into clinical practice.

  • Need for validation studies: While radiomics has shown promise in numerous studies, there is a need for large-scale, multi-center validation studies to confirm these findings. These studies are crucial to demonstrate the reliability and generalizability of radiomics.
  • Despite these challenges, the potential benefits of radiomics in improving patient care and outcomes make it a promising field worth pursuing. With ongoing research and development, it is hoped that these challenges can be overcome, paving the way for the widespread adoption of radiomics in clinical practice.

Standardization and validation

Standardization

Standardization in radiomics is of paramount importance for obtaining reliable results. This involves the harmonization of imaging protocols and analysis pipelines.
The field of radiomics extracts many quantitative features from medical images. However, the results can vary significantly based on the imaging protocol and analysis pipeline used. Therefore, standardization is crucial to ensure the reliability and reproducibility of results. Standardizing imaging protocols helps to minimize variations in image acquisition, such as differences in scanner settings or patient positioning. Similarly, standardizing analysis pipelines ensures that the same procedures are followed when extracting and processing radiomic features.
By adhering to standardized protocols and pipelines, clinicians can have greater confidence in the radiomics results, leading to more accurate diagnoses and treatment plans.

Validation

Validation is crucial in radiomics to confirm that models produce accurate and reliable results across different patient groups. Models developed using specific datasets must undergo validation in diverse populations to ensure their applicability. Without proper validation, a model performing well in one group may falter in another, leading to erroneous predictions and potentially harmful decisions. Validating in diverse populations ensures model robustness and reliability, irrespective of patient

Ethical considerations and future directions

 Ethical issues

Radiomics, like any other medical technology, comes with its own set of ethical considerations. These primarily revolve around patient privacy, informed consent, and responsible use of data.

  • Patient privacy

Patient privacy is a fundamental ethical consideration in radiomics. Data are derived from medical images and clinical records, which are sensitive personal health information. It’s crucial to ensure that this data is stored, processed, and transmitted securely to protect patient privacy.

  • Informed consent

Informed consent is another critical ethical issue. Patients should be fully informed about how their data will be used, what radiomics involves, and any potential risks or benefits. They should have the right to consent to or decline their participation in these types of studies.

  • Responsible use of radiomics data
    The responsible includes using the data for its intended purpose, ensuring data quality and integrity, and avoiding misuse of the data.