Program Information
Knowledge-Based Automatic Pre-Delivery Plan Verification Using Isolation Forest
S Liu1*, Y Fu1 , T Mazur1 , H Li1 , D Yang1 , (1) Washington University in St. Louis, St. Louis, MO
Presentations
TU-FG-702-12 (Tuesday, August 1, 2017) 1:45 PM - 3:45 PM Room: 702
Purpose: To detect major accidents and prevent catastrophic errors, an advanced knowledge-based anomaly detection algorithm based on isolation forest technique, was developed to automatically verify treatment plan parameters right prior to treatment deliveries.
Methods: A software system, i.e. APDV (Automatic Pre-Delivery Verification), was designed to continuously monitor new DICOM plan files on the TMS computer at the treatment console. External beam treatment plan data of the past 8 years at authors’ institution were obtained from Mosaiq, processed and used to train the isolation forest model (namely iForest). When a new plan to be delivered is detected, the consistencies of plan parameters are checked using iForest, a knowledge-based anomaly detection algorithm. By grouping the new treatment plan with the processed plans with the same treatment site, technique and modality, iForest provides a degree of anomaly (anomaly score) by sorting each plan parameter according to their path lengths on the trained trees in the iForest.
Results: 8520 previous plan data were processed. Half were randomly selected for training and the other half were used for testing. The studied parameters were MU, prescription dose, MU/cGy, number of fractions/beams/segments, and plan complexities. Errors were artificially and randomly simulated using boxplot with mild outliers (1.5 IQR). The error detection rate was 99.3%, and the specificity was 94.7%. Average computational time to check one new plan was 2.3 seconds.
Conclusion: This work has shown feasibility to use the iForest method for plan parameter outlier detection. The proposed method has shown high sensitivity and specificity, and have a linear computational time complexity and a low computer memory requirement for high dimensional data.
Funding Support, Disclosures, and Conflict of Interest: Funding: AHRQ R01-HS022888 No conflict of interest Disclosures: Authors have technology licensing fee from Viewray
Contact Email: