An Efficient Tool for Structure Label Harmonization

Abstract

The aim of the study is to develop an efficient solution to addressing the inconsistency of structure labels in treatment plans. Significant efforts have been made in recent years to develop standards for structure labels and other clinical data captured in radiation therapy (RT). While adoption of these standards will enable standardized structure labels in future clinical practices, we need an efficient way of transforming existing clinical cases so that the structure labels all conform to proposed standards. This capability is important for integrating large datasets of RT data to support efforts of rapid learning and precision medicine. It is particularly important when multiple sites in an organ are treated at various time intervals, for example, in repeat radiosurgeries involving multiple brain lesions. We propose a rule-based method for structure label harmonization. Each rule specifies the standard name for an existing structure label that satisfies certain character patterns. These character patterns are described in a special language called regular expressions. The system starts by preprocessing the structure labels, which replaces special characters, separates words by capital letters, and removes any duplicated structure labels. After preprocessing, each structure will have one or multiple well-formed structure labels. By using an algorithm similar to human perception, the system finds similar patterns among the labels from the same structure. This process includes identifying planning keywords, ordering priority by physicians' signature, extracting common strings, and finding acronyms. It then generates a regular expression that forms all the rules. The system will also perform a crosscheck in the generated regular expressions to prevent mismatch between structure labels having similar patterns. Then, it creates a dictionary specifying for each structure, the standard structure name proposed by AAPM TG263 and its corresponding regular expressions. In this study, we used a training set of 104 prostate cases and 102 brain-SRS cases to generate a dictionary. The regular expressions in the dictionary covered all OARs (Organ-at-risk) and non-planning structures in the training set. As an initial validation we applied the regular expressions to a validation set of 100 prostate test cases. The results were encouraging. Ninety-one (91) cases had all structures correctly identified and relabeled. The remaining 9 cases had one or two missing/misclassified structures. These structure labels were due to rare patterns that did not appear in the training cases. Further work is under way to incorporate a learning capability to improve generalization incrementally. This study presents a rule-based approach to harmonize structure labels in existing clinical treatment plans. Preliminary tests demonstrate the feasibility of this system to efficiently and automatically harmonize structure labels across large datasets.