An automated treatment plan alert system to safeguard cancer treatments in radiation therapy.

ABSTRACT

In radiation oncology, the intricate process of delivering radiation to a patient is detailed by the patient's treatment plan, which is data describing the geometry, construction and strength of the radiation machine and the radiation beam it emits. The patient's life depends upon the accuracy of the treatment plan, which is left in the hands of the vendor-specific software automatically generating the plan after an initial patient consultation and planning with a medical professional. However, corrupted and erroneous treatment plan data have previously resulted in severe patient harm when errors go undetected and radiation proceeds. The aim of this paper is to develop an automatic error-checking system to prevent the accidental delivery of radiation treatment to an area of the human body (i.e., the treatment site) that differs from the plan's documented intended site. To this end, we develop a method for structuring treatment plan data in order to feed machine-learning (ML) classifiers and predict a plan's treatment site. In practice, a warning may be raised if the prediction disagrees with the documented intended site. The contribution of this paper is in the strategic structuring of the complex, intricate, and nonuniform data of modern treatment planning and from multiple vendors in order to easily train ML algorithms. A three-step process utilizing up- and down-sampling and dimension reduction, the method we develop in this paper reduces the thousands of parameters comprising a single treatment plan to a single two-dimensional heat map that is independent of the specific vendor or construction of the machine used for treatment. Our heat-map structure lends itself well to feed well-established ML algorithms, and we train-test random forest, softmax, k-nearest neighbors, shallow neural network, and support vector machine using real clinical treatment plans from several hospitals in the United States. The paper demonstrates that the proposed method characterizes treatment sites so well that ML classifiers may predict head-neck, breast, and prostate treatment sites with an accuracy of about 94%. The proposed method is the first step towards a thorough, fully automated error-checking system in radiation therapy.