Editorial: Machine learning in radiation oncology

Wei Zhao, Ye Zhang, Jia Wu, Xiaomeng Li, Yuming JiangFrontiers in Oncology 2022 · DOI · OpenAlex

Abstract

Machine learning (ML) excels in learning complex relationships and incorporating existing prior knowledge into an inference model. The model is able to recognize complex patterns in medical data using either human-engineered features or deep neural network representations. Given the nature of heavy data involvement of ML and a large amount of existing routine data, such as computed tomography images, organs and targets contours, treatment plans, and prognosis records, radiation oncology is one of the best fields for ML algorithms. ML has the potential to benefit the whole workflow of radiotherapy: from patient modeling, image segmentation, treatment planning, patient setup, and prognosis analysis. Indeed, the Food and Drug Administration has cleared numerous ML-based technologies with regard to various radiation oncology aspects and there are having transformative ML-based applications in clinical practice.

Introduction

Radiation oncology is a fundamentally data-intensive discipline. The clinical workflow—spanning from initial diagnosis and tumor identification to the precise calculation of radiation doses—relies on a massive volume of computed tomography (CT) images, MRI scans, complex treatment plans, and longitudinal patient records. Historically, these tasks have been manual and labor-intensive, leading to significant human variability and potential bottlenecks in patient care.

The core challenge in modern radiotherapy is effectively harnessing this "big data" to optimize the pipeline. This editorial synthesizes a collection of 13 research papers that explore how machine learning (ML) can automate these tasks, reduce clinician workload, minimize errors in target delineation, and provide personalized predictions for treatment response.

Methodological Frameworks

The featured studies employ a diverse array of technical strategies to address different stages of the radiotherapy workflow:

  • Deep Learning (DL) Architectures: Convolutional Neural Networks (CNNs), including ResNet18 and modified DenseNet architectures, are utilized for automated image segmentation and classification tasks.
  • Transfer Learning: This technique allows models pre-trained on large, general datasets to be fine-tuned for specialized clinical tasks, such as differentiating rare brain tumors like germinomas from more common gliomas.
  • Radiomics and Dosiomics: Researchers extract high-dimensional features from medical images (radiomics) and spatial dose distributions (dosiomics) to build predictive models for treatment outcomes and potential adverse side effects.
  • Multi-task Learning: Some frameworks simultaneously perform tumor masking and dose optimization, significantly accelerating the treatment planning process.
  • Synthetic Imaging: CNNs are being used to generate "synthetic" functional data, such as lung perfusion images, from standard anatomical CT scans, potentially reducing the need for invasive or expensive diagnostic procedures.

Data Foundations

The effectiveness of these ML models is grounded in a variety of clinical and experimental datasets, reflecting the breadth of the field:

  • Thoracic CT Scans: Essential for training models in automated "Organs-at-Risk" (OAR) delineation.
  • Multiparametric MRI: Sequences like T1-contrast (T1C) are used for brain tumor differentiation and pretreatment planning for glioblastoma.
  • Contrast-Enhanced CT: Used specifically for identifying colorectal liver metastases.
  • Time-Resolved Microscopy: Combined with phase-contrast and Cytotox Red images to model radiobiological responses at the cellular level.
  • Clinical Outcome Records: These longitudinal records are vital for training models to predict local responses to Gamma knife radiosurgery and the likelihood of adverse radiation effects (ARE).

Clinical Insights and Innovations

Several critical themes emerge from this collection of research that signal a shift in clinical practice:

  1. Reaching Clinical Parity: A significant finding is that automated DL-based segmentation for certain thoracic organs shows no statistically significant difference from manual segmentation performed by experts. This indicates that ML is reaching a maturity level where it can reliably handle the "heavy lifting" of contouring.
  2. The Rise of Dosiomics: While radiomics focuses on anatomy, the rise of dosiomics—analyzing the 3D distribution of the radiation dose itself—marks a shift toward a more holistic analysis of treatment toxicity and efficacy.
  3. Data Quality and Homogeneity: Research suggests that a homogeneous training dataset (with consistent imaging and planning parameters) can outperform larger but more heterogeneous datasets for dose prediction. This underscores the importance of standardized data collection in clinical AI.
  4. Unlocking Hidden Information: The ability to synthesize functional imaging from anatomical data suggests that ML can extract latent information from standard scans, streamlining the diagnostic process and reducing patient burden.

Quantitative Results

The studies demonstrate impressive quantitative results that validate the potential of these technologies:

  • Diagnostic Precision: One study achieved a mean Area Under the Curve (AUC) of 0.88 when using a fine-tuned ResNet18 to differentiate germinomas from gliomas.
  • Superiority over Atlas-Based Methods: DL-based segmentation for nasopharyngeal carcinoma significantly outperformed traditional atlas-based methods, particularly in small and anatomically complex volumes.
  • Efficiency Gains: Multi-task learning frameworks have been shown to drastically shorten the treatment planning procedure by integrating segmentation and dose prediction into a single computational step.

Future Outlook and Challenges

Despite these advancements, several hurdles remain for universal clinical adoption. The "black box" nature of deep learning models necessitates rigorous clinical justification and validation. Furthermore, model performance is highly sensitive to training data distribution; a model optimized for one institution may require significant recalibration to generalize to another.

Future efforts must focus on developing robust decision-support tools that integrate seamlessly into existing clinical workflows, ensuring that ML serves as a reliable partner to clinicians rather than a standalone "silo" of information. The goal is to move beyond theoretical research into transformative clinical tools that enhance both precision and efficiency in patient care.

Summary updated

Abstract

Machine learning (ML) excels in learning complex relationships and incorporating existing prior knowledge into an inference model. The model is able to recognize complex patterns in medical data using either human-engineered features or deep neural network representations. Given the nature of heavy data involvement of ML and a large amount of existing routine data, such as computed tomography images, organs and targets contours, treatment plans, and prognosis records, radiation oncology is one of the best fields for ML algorithms. ML has the potential to benefit the whole workflow of radiotherapy: from patient modeling, image segmentation, treatment planning, patient setup, and prognosis analysis. Indeed, the Food and Drug Administration has cleared numerous ML-based technologies with regard to various radiation oncology aspects and there are having transformative ML-based applications in clinical practice.
BibTeX
@inproceedings{zhao2022editorial,
  author = {Wei Zhao and Ye Zhang and Jia Wu and Xiaomeng Li and Yuming Jiang},
  title = {{Editorial: Machine learning in radiation oncology}},
  booktitle = {Frontiers in Oncology},
  year = {2022},
  doi = {10.3389/fonc.2022.1032858},
}
Figure