Page 107 - ARNM-3-2

Advances in Radiotherapy
& Nuclear Medicine Artificial intelligence in radiotherapy

treatment strategies. AI offers predictive algorithms to instance, questions such as which algorithm should be
1
predict treatment-related toxicities. This approach not used in specific scenarios, which techniques should be
5,6
only improves patients’ quality of life but also prevents employed to extract radiomic features, and how many
unnecessary healthcare expenditures. cases are required for algorithm training to achieve
accurate results remain unanswered and need to be
Patient follow-up after treatment is another area where
the use of AI is expanding. Algorithms monitor treatment standardized. Training AI systems, standardizing them,
responses, enabling clinicians to intervene early. Moreover, and making them suitable for routine clinical practice
AI-based algorithms developed with appropriate baseline require large volumes of patients’ data. However, patient’s
data can predict patients’ oncological outcomes, allowing confidentiality and data security are critical concerns. The
for more radical treatments or avoiding unnecessary risk of unauthorized access poses a significant challenge
therapies. 7,8 for healthcare providers and researchers. Therefore,
anonymization and secure storage of data are essential.
Mobile applications are also gaining importance. Integrating AI-based systems into clinical applications
Remote monitoring systems are effectively used to assess and implementing them in routine practice is a lengthy
post-treatment symptoms and quality of life. Mobile and complex process. Without sufficient evidence on
15
applications can be leveraged to provide patient education, the effectiveness and reliability of these technologies,
thereby improving treatment adherence. 9 pre-mature implementation in clinical settings could
In radiation oncology, an increasing amount of lead to incorrect treatments and unexpected toxicities,
heterogeneous data is accumulated, including clinical resulting in a loss of trust among clinicians and patients.
information, planning computed tomography scans, and In addition, aligning existing clinical workflows with AI
dosimetric data. Machine learning plays an increasingly systems can be time-consuming and costly.
significant role in RT processes due to its ability to analyze AI models can also be affected by biases present in
large and heterogeneous datasets to make predictions. the training datasets. For example, algorithms trained on
Specifically, algorithms such as support vector machines, datasets lacking in demographic diversity may produce
random forests, and k-nearest neighbors are used to analyze inaccurate results for certain patient groups. This could
patient’s data for toxicity prediction and the evaluation of lead to disparities and inaccuracies in treatment outcomes.
treatment outcomes. 7-10 Thus, ensuring data diversity and ethical considerations

DL models have enabled substantial advancements during model development is crucial. 16
in RT planning and image processing. Convolutional Ensuring that AI models provide consistent results
neural networks provide high accuracy, particularly in across diverse patient populations and clinical conditions
tumor segmentation and OAR delineation processes in is essential for their clinical application. However,
medical imaging. 11,12 Through the radiomics approach, the generalizability of these models cannot always be
advanced features imperceptible to the human eye can be guaranteed. Models trained on small or imbalanced
extracted from radiological images, offering insights into datasets may perform poorly in different patient
tumor biology. This method facilitates the integration of populations, highlighting the need for thorough validation
imaging data with genetic, proteomic, and metabolomic and optimization of these systems. The inclusion of AI in
data, playing a critical role in personalized medicine clinical decision-making processes also introduces ethical
applications. In addition, dosiomics, which uses radiomic and legal challenges. For instance, if an AI system makes an
1
tools to characterize RT dose heterogeneity, provides more error, who should bear the responsibility – the clinician or
comprehensive information compared to traditional dose- the technology provider? As AI continues to be increasingly
volume histograms. AI enables the analysis of this high- utilized in healthcare, it is imperative to develop unbiased,
dimensional data to yield clinically meaningful outcomes. data-driven algorithms that are frequently monitored and
By automating RT planning, AI offers significant time updated.
and resource efficiency. Algorithms used in areas, such as While AI is unlikely to entirely replace clinical
automatic segmentation and dose distribution optimization judgment, it can aid clinicians in making better decisions.
accelerate workflows while ensuring standardization and Unlike human’s decision-making processes, AI’s decisions
enhancing treatment quality. 13,14 and judgments are systematic, operating within defined
Although AI provides innovations and conveniences algorithms. However, in the absence of an effective legal
in radiation oncology, there are challenges and framework, the responsibility for AI-related decisions
limitations associated with its adoption and integration. currently lies with those who design and use these
Standardization is still lacking in many areas. For systems. 17

Volume 3 Issue 2 (2025) 99 doi: 10.36922/arnm.8429