Treatment plan approximation based on patient database

Radiation therapy is a form of cancer treatment in which ionizing radiation is used to damage cancerous tissue. The patient is treated over multiple occasions (called fractions), during which a carefully designed treatment plan directs the delivery of the radiation. The radiation plan is computed prior to the first treatment in order to administer the maximum dose possible to the target and the minimum dose possible to any nearby organs at risk. At the beginning of the treatment planning process, the patient's anatomy is acquired by a medical imaging modality. On this anatomical image, contour lines are manually or semi-automatically drawn around all structures of interest. The optimized radiation treatment plan can be created using the CT scan, the contours, and the prescription, and usually takes a few hours.

Historically, the initial plan was used througout the fractions, which introduced a geometric uncertainty as the treatment proceeded. However, as the fractions take place over an extended period of time, numerous changes occur between the treatments. One such change is varying patient setup that is compensated by image guidance, when the measured patient geometry is matched to the planning geometry in a rigid fashion. Other changes also occur, such as tumor shrinkage/expansion, or movement of the surrounding organs at risk (e,g, in prostate cases the prostate moves due to bladder and rectum filling), which however, cannot be compensated rigidly. These "deformable" deviations need to be handled by modifying the initial treatment plan. Creating a new adaptive daily plan is quite laborious, so using the current manual workflows it is not feasible to perform as often as it would be needed to reach optimal treatment quality.

Great advancements have been made to speed up the most time consuming steps in treatment planning. Inverse treatment plan optimization creates the most efficient beam configuration for a treatment plan, and automatic segmentation (contouring) delineates the structures of interest without or with minimal user interaction. Automatic segmentation however, is still far from being able to use routinely in a clinic in a reliable way, so other approaches must be explored to facilitate automatic treatment planning.

Clinics store a vast amount of patient data, including a huge treatment plan database of past treatments. These plan databases can be potentially exploited by comparing the database contents to the current situation. We hypothesize that approximating the current treatment plan by finding and using the most similar one in the plan database will significantly speed up the treatment creation process, both in case of initial treatment planning and daily plan creation.