A RADIOTHERAPY PLAN SELECTOR USING CASE-BASED REASONING

Abstract

Developing a head and neck cancer treatment plan for a candidate of Intensity Modulated Radiation Therapy (IMRT) requires extensive domain knowledge and subjective experience. Therefore, it takes a cancer treatment team at least 2 to 3 days to develop such a plan from scratch. Many times the team may not use a reference plan. Sometimes, to reduce the amount of time taken to generate each treatment plan, these experts recall a patient, whose plan they recently prepared, and who had similar symptoms as the candidate. Using this recalled patient's plan as the starting point, the cancer treatment team modifies it based on the differences in the symptoms of the new candidate and those of the reference patient record. The resultant plan after modification is presented as the new treatment plan for the oncologist to evaluate its suitability for treatment of the candidate. This approach is heavily dependent on the team's choice of the reference patient record. Choosing a starting treatment plan where the patient's symptoms are not the closest to the new candidate implies that more time will be spent modifying the plan than is necessary and the resultant treatment plan may not be the best achievable under the same circumstances given a better starting plan. Therefore, the team's bias in choosing the starting plan may affect the quality of treatment plan that is finally produced for the candidate. This thesis proposes a system that behaves like an un-biased radiotherapy expert - following a similar process and standards as the human experts and which searches the entire IMRT patient database and returns the record (with patient symptoms and treatment plan) for a patient whose symptoms are most similar to the candidate's symptoms. It takes in the new candidate's information (from diagnosis, scans of the tumor and interviews with the candidate), searches the database and prints out a patient record showing another patient's treatment plan as the suggested starting point for generating the new plan. The system uses Case-Based Reasoning (CBR) because it mimics the experts' approach since it makes use of previous successes and shuns reasoning that has failed in the past. This occurs by considering only treatment plans that have been implemented successfully on patients in the hospital archive. For this thesis, CBR is applied using fuzzy IF-THEN rules to search the patient database. Fuzzy logic is used because it can handle imprecise expressions commonly used in natural language to determine the appropriate weight of the patient attributes in the search process. Filtering of patient records based on parameter value ranges is also used to reduce the number of records that have to be compared. The system code developed for this thesis was prepared in Java and C Language Integrated Production System (CLIPS) using the Java Expert System Shell (JESS). This system is part of a bigger expert system that is being prepared by the Intelligent Systems Applications Center (ISAC) for Thomas Jefferson University Hospital, expected to generate a radiotherapy plan for a patient designated for IMRT treatment. Initial results from the developed prototype prove the viability of selecting similar patients using CBR. It is important to note that the overall objective of the project is to build a system that effectively aids decision support by the IMRT team when generating a new treatment plan and not to replace them. The team is expected to use the generated plan as a starting point in determining a new treatment plan. If the generated plan is sufficient, the oncologist and their team will have to check this plan (in their various capacities) against expected standards for quality control before passing it on for implementation. This will save them time in planning and allow them to focus more on the patient's needs hence a higher quality of life for the patient after treatment.