Cancer treatment planning

Elekta’s Leksell Gamma Knife® (stereotactic radiosurgery)

The introduction of Elekta’s Leksell Gamma Knife® PERFEXION™ took the field of stereotactic radiosurgery to a new level. Unlimited access to cranial volume and full automation approach open a new research area for the researchers who have been already working on the previous Lekshell Gamma Knife unit for years. Although this method is primarily designed to be performed on brain tumors, today, it is considered an effective treatment for several other conditions, including arteriovenous malformations and pituitary tumors.

The new capabilities of the PERFEXION™ unit and its successors require new methods of treatment plan design. In previous Gamma Knife® designs, the collimator (the component that adjusts the shape/location of the radiation delivery) had to be manually adjusted, rendering complex treatments too labor intensive for clinical viability. Because the movement of the collimator in the new PERFEXION™ unit is now automated, complex treatment plans that can very tightly conform to the targeted treatment area can be delivered in clinical settings.

To deliver a high quality treatment, we select collimator positions based on optimization methods. The optimization algorithms are designed to deliver an appropriate amount of dose to the target area while simultaneously avoiding sensitive healthy tissues, thereby leaving the patient with a high quality of life and fewer side effects after treatment. These optimization approaches also allow us to explore new forms radiosurgery treatment, including dynamic treatments where the patient is moved continuously through the radiation isocentre, and dose-painting, a highly heterogeneous form of dose specification to treat complex areas.

Automated tumour board decisions

Multidisciplinary tumour boards (MDTs) are treatment planning forums where a group of specialists meet and discuss the diagnosis and management of patients with cancer, and how to proceed with treatment: surgery, local radiation treatment, or non-local treatment. At our partner institution, Princess Margaret Cancer Centre (Toronto, Canada), MDTs regularly take place for most cancers, including a specialized weekly meeting for lung cancer metastases patients attended by thoracic surgeons, radiation oncologists, and radiologists. At the weekly meetings of this tumor board, all lung metastases cases are discussed, and not only specific complex cases, as is the case in tumor boards in other anatomic sites.

Due to the high volume of cases reviewed by MDTs, and the time and costs associated with convening such a board with a large number of participants, routine cases may take up disproportionate institutional resource utilization, leaving limited resources for more complex cases. Changes in participants across multiple sessions of the board can lead to a lack of consistency, and the quality of patient information available to tumor board participants is a major contributor to the efficacy of decision-making; therefore, missing, or inaccurate patient data can hamper MDTs.

We develop automated machine learning tools to predict treatment directions to provide MDTs with recommended decisions, allowing straightforward cases to be decided quickly, leaving more time for discussion of complex cases. Our predictive tool provides accurate decisions even when patient data is imperfect, and also improves the consistency of decisions across different MDTs. This tool can additionally be implemented in hospitals that treat cancer patients without having the same breadth of expertise available at large teaching and research hospitals.

Automated quality assurance

The process of validating a radiotherapy treatment plan requires a quality assurance review by an expert. If the plan is deemed acceptable, it proceeds to treatment. Otherwise, it is returned to a dosimetrist to revise the treatment to meet standards for the particular clinic. This review is time-consuming, taking as much as tens of thousands of man-hours annually at large treatment centers, and may be subject to human errors.

The goal of this research is to automate the QA process using machine learning to learn clinic treatment standards and flag treatments that do not meet those standards. There can be little tolerance for incorrectly identifying an erroneous plan, as there may significant health implications for the patient. Notably, a plan that does not conform to standards may not be erroneous; it may be part of trial of new treatments or the patient may require unusual treatment. Thus, any flagged plan must be manually reviewed.

The challenge in correctly classifying plans as acceptable or not is twofold. First, radiotherapy QA datasets are extremely high-dimensional due to the number of features required to represent dose distributions. Second, the datasets are highly imbalanced in that most stored plans are acceptable, as unacceptable plans are generally overwritten with improved plans. We develop classification and anomaly detection tools to identify unacceptable plans with a low false negative rate, and interpretation methods to communicate to dosimetrists how to improve unacceptable plans.