A recent perspective, published in Nature Reviews Clinical Oncology, reviews the opportunities and challenges of implementing artificial intelligence (AI) in radiation oncology. The authors explore how AI might increase the efficiency, accuracy and quality of radiation therapy, which will improve how cancer care is delivered within resource-limited environments.
A pillar of cancer treatment
Radiation therapy is a critical pillar of cancer treatment and oncologists use it for ~50% of patients. Nevertheless, estimates indicate that millions of patients currently lack access to this vital treatment modality. Despite technological advancements, the radiation therapy workflow is still very time-consuming, requiring the manual input of a diverse team of healthcare professionals. In recent years, there has been shortages in the radiation oncology workforce. In addition, knowledge gaps and experience between adequately and under-resourced systems across the globe has resulted in inequalities in cancer care.
AI in the workflow
AI is transforming the medical field and has the potential to address many of the challenges faced in radiation therapy. As a result, AI could improve the availability and quality of cancer treatment worldwide. The radiation therapy workflow involves a multitude of complex tasks. These include tumour and organ segmentation, dose optimisation, outcome prediction and quality assurance. Deep learning allows for the aggregation of these different data streams, which can improve clinical decision-making. Consequently, the researchers in this Perspective suggest how to apply various AI algorithms to each task of the radiation therapy workflow (summarised below).
Initial treatment decision-making
- Patient evaluation: The workflow begins with patient intake and evaluation, which usually involves a consultation by the radiation oncologist. After the consultation, the oncologist will recommend a treatment plan. AI in this stage could automatically extract key clinically actionable features from clinical, genomic and imaging data, which will be important in building decision support tools for clinicians.
- Dose prescription: The radiation oncologist determines the prescribed dose of radiation prior to treatment planning. This is in accordance with nationally accepted standards and evidence from clinical trials. AI platforms could enable personalisation of radiotherapy by predicting the radiation sensitivity of the tumour and the optimal dose.
Treatment planning and preparation
- Treatment simulation: Prior to treatment planning, simulation appointments take place where medical images are acquired to help inform the treatment plan. This process can be very complex. AI could offer solutions for potential challenges based on prior knowledge of patients’ anatomy. It could also reduce exposure of radiation, enhance the quality of images, suppress artefacts and enable more accurate image registration.
- Image segmentation and treatment planning: Manual segmentation of the primary tumour and affected lymph nodes is one of the most crucial but time-consuming tasks performed by radiation oncologists. The accuracy of this task can directly affect outcomes. AI has the potential to automate segmentation approaches. This could increase efficiency, reproducibility and quality of radiation treatment. Another key component in treatment planning is aiming to maximise the dose delivered to the tumour while sparing surrounding organs. AI tools for automating treatment planning have shown promise in predicting optimal dose distribution.
Pre-treatment review and verification
- After the radiation oncologist’s approval of the treatment plan, medical physicists perform plan checks and other QA checks. Experts have developed AI tools to minimise the need for repetitive, time-consuming manual measurements and to improve the efficiency of some QA activities.
Treatment setup and delivery
- Scheduling: Patients receiving radiation therapy have to attend several appointments. Each appointment will have a different waiting time and duration. AI could identify the most important factors contributing to waiting time durations and predict waiting times. This would optimise clinic flow and efficiency.
- Image guidance and motion management: A key part of radiotherapy delivery is setting the patient up in the same position that was used to create the treatment plan. Experts have applied AI to improve the image quality of integrated cone beam CT (CBCT) to allow more accurate positioning of patients for treatment.
- Adaptive treatment: Changes in patient anatomy prior to treatment and throughout can warrant re-planning. This involves creating a new treatment plan based upon up-to-date images of the patient’s anatomy. Radiation oncologists currently determine when anatomical changes are clinically relevant. AI could predict which patients require adaptation of treatment and the ideal time point at which it should occur.
Completion of treatment
- Response assessment and follow-up care: The Response Evaluation Criteria in Solid Tumours (RECIST) is a widely adopted system for evaluating treatment response. AI algorithms have the potential to provide more detailed information on tumour responses throughout the course of treatment. Researchers have also found that these algorithms can detect early changes in some cancers.
- Toxicity prediction and management: The unpredictable occurrence and severity of adverse effects complicate the management of toxicities in patients. AI could build more robust predictive models which could help in clinical decision support for both anticipatory management and secondary prevention of toxicities.
Challenges
The implementation of AI tools into the clinic will require upfront investment of both time and resources. There must be efforts to understand the utility and limitations of these tools alongside efforts in redesigning current clinical workflows. In addition, establishing trust of these AI systems is critical, given the ‘black box’ nature of many algorithms.
Tasks that could have a big impact on patient treatment which AI could complete or assist in will present a particular challenge in clinical implementation. Governing of algorithm-based decision-making has yet to be fully developed. While AI has the potential to reduce medical errors, it is expected to alter the legal landscape around clinical liabilities and responsibilities. There are also growing concerns towards biases within AI and the potential unethical approaches that could be developed.
Conclusion
Evidently, as AI becomes integrated into clinics, the roles of staff members will need to be redefined. Experts estimate that over half of all cancer patients live within low-income or middle-income countries. The availability of AI tools will help with workforce and equipment shortages and will drastically improve efficiency and throughput in radiation therapy.
Image credit: Kishore Kumar – canva.com
