Integrating a Tool to Automatically Determine Necessity of Online Adaptive Replanning

Abstract

As online adaptive replanning (OLAR) is labor-intensive and time-consuming, it's desirable to determine when OLAR is necessary before OLAR is initiated. We have previously reported a novel method to automatically determine the necessity of OLAR using machine leaning algorithms based on the structural similarity maps (SSIM) and wavelet texture maps (WMT) extracted from the daily MRI during MR-guided adaptive radiation therapy (MRgART). This study aims to integrate this method into a commercial software platform that has been used during our routine MRgART. The method of automatically determining the necessity of OLAR based on daily MRI was implemented and integrated into the software platform through a specifically developed workflow. The obtained workflow was tested using 25 daily MRI sets acquired from 5 patients with pancreatic cancer in the following procedure: 1) rigidly registering the daily and reference MRIs, 2) identifying the region enclosed by the 50-100% iso-dose surfaces on the daily MRI by transferring the iso-dose surfaces from the reference to the daily MRIs, 3) launching our in-house codes to calculate significant changes in textures extracted from SSIM and WMT maps, 4) inputting the feature values into the pre-trained classifier models for SSIM and WMT, and 5) outputting results considering the WMT based prediction as the primary indicator and the SSIM-based as the secondary (validation) indicator on whether OLAR is needed for the daily MRI. The execution of the developed workflow was fast and can be used to streamline the process. It provides the ability to scroll through the images for better decision making while providing quantitative prediction within 30-38 seconds. Eighty percent of the daily MRIs required OLAR. The SSIM map displayed was able to successfully captured the areas of similarity between the reference and daily MRIs and the WMT prediction agreed with the prediction class. The integration of the prediction method for automatically determining the necessity of OLAR based on two independent machine learning classifiers into a commercially available software is feasible and can be used to streamline the process of MRgART. With larger verification studies, this workflow-based tool may be developed into a generalized tool that assist in OLAR using different platforms.