The Use of Machine Learning and Artificial Intelligence in Radiation Oncology

Numerous computational methods have been developed to improve radiotherapy dosimetric tasks. Such popular approaches, such as machine learning (ML) and artificial intelligence (AI) integrate big data and enhance the capabilities of a single dosimetrists whilst reducing the planning time. Although these techniques have enhanced the quality of patient care, workflow efficiency is dictated by their ability to mitigate multiple sources of error. This review examines big data, machine learning (ML) and artificial intelligence (AI) both generally and in the context of radiotherapy. Further, the analysis mirrors a critical review of their current and /or potential future use, and their importance in radiation oncology.

“Big data” refers to large amounts of data in terms of volume, veracity, variety, and velocity (Bibault, et al. 2016). Lustberg et al. (2017) reports a rapid increase in the use of data-intensive diagnostic and imaging modalities as defined in this context. Aligned with this definition, “big data” in radiation oncology is the act of studying a considerable cohort of patients along heterogeneous sets of diagnosis parameters (Bibault, et al., 2016; Lustberg et al., 2017). Such data can elucidate patterns beyond human understanding when analysed using unsupervised machine learning tools (Bibault, et al. 2016).

Kang et al. (2015) defines machine learning (ML) as the process of using statistical models to predict patterns from complex data sets. These methods apply in different areas, such as filtering junk mails, face recognition, handwriting recognition, and speech recognition. ML tools have been used to predict the outcomes of radiation therapy. They include artificial neural networks (ANN), support vector machines (SVM), and logistic regression (LR) (Kang et al., 2015).

Artificial intelligence (AI) is a computerized method of identifying complex numerical relationships within varied sets of observational data (Thompson et al., 2018a). The explosion of AI capabilities has been facilitated by recent developments in cloud-based computing frameworks and computing algorithms. In the context of radiotherapy, AI is the process of integrating or leveraging orthogonal data (such as electronic medical records, imaging, diagnostics, genomics) with dosimetric information to improve radiation dose prescription and treatment modality (Thompson et al., 2018a).

Big data handling requires a planning system that is easy to use so as to support clinical decisions (Bibault, et al. 2016). With such a system in place, the physicist or dosimetrist can obtain a patient’s medical history and administer an achievable treatment plan. In addition, the same system can be used to notify the physician in case a deviation from the predicted norm occurs during treatment.  Most health care institutions use Electronic Health Records (EHRs) to capture vast amounts of data on treatment follow-ups, adverse events, treatment plans, and patient characteristics. As poor data generates poor results, emphasis is placed on the nature and quality of the wealth of information in EHR systems. However, clinical trials are used alongside such systems to collect detailed data and to avoid confounding variables.

Research has focused on the integration of patient-specific big data (volume, veracity, variety, and velocity) into a single model (Bibault, et al. 2016; Lustberg et al., 2017). However, the realization of this vision would require substantial improvements in the current methodological approaches (Bibault, et al. 2016). For example, it will be imperative to integrate biological, dosimetric, and clinical data into a single model, which will be validated across a selected cohort of patients. The field of oncology is shifting from histological and anatomical-based therapies, to the more complex molecular abnormalities. However, the individualization of molecular traits implies that it will become increasingly difficult to achieve sufficient disease characterization with clinical trials (Lustberg et al., 2017). Further, the design of such trials will become unsustainable if methodological issues and financial constraints are not addressed.

Big data efforts in radiotherapy emphasize on the integration of treatment results with patient outcomes, coupled with clinical decision support (Lustberg et al., 2017). The objective is to provide relevant point-of-care recommendations by leveraging on existing sources of data. The role of clinical decision support in radiation oncology is twofold: (i) developing accurate models of tumours using techniques, such as Normal Tissue Complication Probability (NTCP) and Tumour Control Probability (TCP), using dosimetric data, and (ii) balancing dosimetric trade-offs so that assurance agencies, clinicians, and patients can make informed decisions about dose prescriptions and treatment modalities (Kang et al., 2015). 

Most of the ML tools share common overlapping properties and statistical techniques. ANN, SVL, and LR are used in clinical settings to evaluate and develop scoring systems for mortalities arising from chronic health and acute physiology after 24 hours (Kang et al. (2015). In particular, SVM and ANN have been used to analyze general survival rates in radiotherapy; ANN is the most used technique in predictive modelling while deep learning (DL) is gaining momentum in diverse fields (Bibault, 2016). However, they are yet to be deployed for more sophisticated predictions in radiation oncology (Kang et al., 2015). In addition to image processing and localization of tumour motion, research has also evaluated the use of ML in predicting the normal toxicity of tissues (Hao, et al., 2012).

Previously, tissue modelling issues were quantitatively addressed via clinical analysis of normal tissue effects. The flaws reported in such clinical methods include high dose-volume histogram (DVH) thresholds and unavailability of unreliable predictors, which leave out ML as the mainstream predictive modelling framework (Bentzen et al., 2010, cited in Kang et al., 2015). NTCP is an ML-based approach to computing normal tissue response based on fractions of a non-uniform arbitrary organ. This tool employs the Kutcher-Burman power law. While DVH-scalar equivalents parameterize these models, they fail to exploit the benefits of non-dosimetric parameters.

SVM accounts for the bimodal distribution by altering the similarity function of the scoring system. This is achieved by squaring the distance, d, between the remaining data points and the average size of a tumour. Subsequently, the data can be isolated on a linear scale after the transformation. In radiation oncology, SVM selects the data points that determine the linear line (Figure 1). These data points represent a combination of predictors across a selected group of reference patients (Kang et al., 2015).

Figure 1: Quadratic transformation of linearly isolated data in 2-dimensions. The black line indicates the decision boundary of the SVM. The grey and while dots denote malignant and benign tumours, respectively (Kang et al., 2015 p. 1132)

With LR, the combination of predictors is linearly mapped to an S-curve. The ML technique is best suited for unrelated predictors (stage of tumour, sex, age, and so forth) and fewer questions. For example, in Figure 2 below, a 1-dimensional question (question with a single predictor) could be in the form of: “Is the tumour size a reliable predictor for malignancy?” (Kang et al., 2015).

Figure 2: 1-dimension illustration of linearly separable data where the grey and while dots denote malignant and benign tumours, respectively (Kang et al., 2015 p. 1133)

Much emphasis is made on these support vectors, implying that the outcomes of SVM depend on these critical points. In contrast, the decision boundary created by LR is one decision lower than the data set, which renders LR is extremely biased by outliers along the boundary decision (Kang et al., 2015). ML techniques are used to support quality assurance programs in radiation therapy. Programs, such as time-series analyses, error prevention, radiation detection, and treatment are essentially suited for ML (Feng et al., 2018).

The ability of ML algorithms to automatically detect outliers permits physicians to focus explicitly on specific aspects that impact on patient care (Feng et al., 2018). ML is also used in radiotherapy to predict deviations from planning expectations. For instance, Varfalvy et al. (2017, cited in Feng et al., 2018) applied hidden Markov models and relative gamma analysis to classify patients based on deviations from the original treatment plan. The same tools were used to predict the likelihood of patients benefiting from a re-planning arrangement (Varfalvy et al., 2017).

The ML applications described above are important, albeit in in different facets. However, much effort is still needed to make them commercially available (Feng et al., 2018). Going forward, ML has the potential to transform the modalities of patient follow-ups, especially those with definitive therapy. The response of tumour markers and imaging changes (enhancement losses or restricted diffusion) have been evolving after radiation therapy. These dynamics can be monitored in real-time to signal therapeutic efficiency. Therefore, ML-based prediction models will be required to speculate treatment outcomes (Feng et al., 2018a).

 AI pertinent areas that have attracted renewed interest in radiotherapy include quality assurance, optimization of doses, inverse planning, and image segmentation of normal tissues. Related applications include Precision Radiation Oncology Platform (by Oncora Medical), Mirada’s DLCExpert, Medical QuickMatch (by Siris), and DeepMind (Google’s) (Thompson et al., 2018a). AI has been able to minimize the risk of toxicity and to predict tumour control using technologically feasible options. With AI, it is now possible to integrate patient data (such as imaging and HER data) from multiple sources (Thompson et al., 2018) and to tailor the treatment modality. In essence, it has reduced human intervention in treatment intervals and has promoted a paradigm shift towards cloud-based adaptive planning methods (Thompson et al., 2018).

The weak points of radiation oncology lie in transcription processes involving data transfer from one therapeutic stage to another. In particular, human involvement aggravates error at the critical junctions. Weidlich et al. (2018) established that AI was effective in error prevention and enhancing process efficiency when employed in automated data transfer processes. In radiation oncology, ML and AI are expected to solve numerous long-standing challenges and to improve efficiency in workflow management (Thompson et al., 2018a). For example, AI algorithms may be applied in machine-level data to minimize technical failures and machine downtime. In tandem, the work efficiency will improve and patient satisfaction indices will rise (Thompson et al., 2018a).

With the advent of AI, ML, and big data integration, the importance of traditional tools such as contouring will wane in the near future. Such algorithms will revolutionize different aspects of radiotherapy, such as patterns of care, training requirements, and resource utilization (Weidlich et al., 2018). The advent of these adaptive machine-human interfaces has positively transformed radiotherapy. However, a future fueled by these tools will be marred with multiple challenges and caveats. For example, AI can cause multiple hazards to patients if used inadvertently (Thompson et al., 2018). Therefore, radiation oncologists must be adequately trained to ensure patient safety. Ultimately, ML and AI will be part of the future of radiation oncology. At this juncture, the opportunity cost of dismissing them is unstoppable.

References

• Bentzen, S.M., et al. (2010). Quantitative Analyses of Normal Tissue Effects in the Clinic (QUANTEC): An introduction to the scientific issues. International Journal of Radiation Oncology, Biology*Physics, 76(3), 03-09. doi:10.1016/s0167-8140(11)71723-6
• Bibault, J., Giraud, P., & Burgun, A. (2016). Big Data and machine learning in radiation oncology: State of the art and future prospects. Cancer Letters, 382(1), 110-117. doi:10.1016/j.canlet.2016.05.033
• Feng, M., Valdes, G., Dixit, N., & Solberg, T. D. (2018). Machine Learning in Radiation Oncology: Opportunities, Requirements, and Needs. Frontiers in Oncology, 8, 110. doi:10.3389/fonc.2018.00110
• Hao H.Z., et al. (2012). Machine learning applications in radiation therapy. In: Machine Learning Algorithms for Problem Solving in Computational Applications: Intelligent Techniques by Siddhivinayak K. Hershey, PA: IGI Global, p.59-84.
• Kang, J., Schwartz, R., Flickinger, J., & Beriwal, S. (2015). Machine Learning Approaches for Predicting Radiation Therapy Outcomes: A Clinicians Perspective. International Journal of Radiation Oncology*Biology*Physics, 93(5), 1127-1135. doi:10.1016/j.ijrobp.2015.07.2286
• Lustberg, T., van Soest, J., Jockeys, A., Deist, T., van Wijk, Y., Walsh, S., … Dekker, A. (2017). Big Data in radiation therapy: challenges and opportunities. The British Journal of Radiology, 90(1069), 20160689. doi:10.1259/bjr.20160689
• Thompson, R. F., Valdes, G., Fuller, C. D., Carpenter, C. M., Morin, O., Aneja, S. . ., Thomas, C. R. (2018a). Artificial intelligence in radiation oncology: A specialty-wide disruptive transformation? Radiotherapy and Oncology. doi:10.1016/j.radonc.2018.05.030
• Thompson, R. F., et al. (2018b). The Future of Artificial Intelligence in Radiation Oncology. International Journal of Radiation Oncology, 102(02), 247-248. doi:https://doi.org/10.1016/j.ijrobp.2018.05.072
• Varfalvy, N., Piron, O., Cyr M.F., Dagnault, A., & Archambault, L. (2017). Classification of changes occurring in lung patient during radiotherapy using relative γ analysis and hidden Markov models. Medical Physics, 44:5043–50. doi:10.1002/mp.12488
• Weidlich, V., & Weidlich, G. A. (April 13, 2018). Artificial Intelligence in Medicine and Radiation Oncology. Cureus, 10(04), 01- 07/e2475. doi:10.7759/cureus.2475