Personalized auto-segmentation for magnetic resonance imaging-guided adaptive radiotherapy of large brain metastases.

Abstract

Magnetic resonance-guided adaptive radiotherapy (MRgART) may improve the efficacy of large brain metastases (BMs)(≥2 cm), whereas the workflow requires optimized. This study develops a two-stage, personalized deep learning auto-segmentation (DLAS) model to assist online delineation of large BMs. Multi-sequences images from 177 BMs were trained to develop the basic DLAS model. Then, 741 daily online MR images of 20 large BMs from a prospective trial were collected for developing a personalized model. The dice similarity coefficient (DSC) was evaluated across three methods: basic model, rigid registration and personalized model, at intervals of every five fractions. The accuracy and efficiency were compared between manual delineation (MD) and DLAS assistant delineation (DLAS-AD) in 8 patients who underwent contrast T1 re-scan during MRgART. The personalized DLAS model demonstrated significantly better performance compared to the basic model and rigid registration during the last fraction of MRgART (when the tumor volume achieved a significant reduction). The mean DSC for basic model vs. rigid registration vs. personalized model were 0.86 (p = 0.01) vs. 0.88 (p = 0.05) vs. 0.90, respectively. The DLAS-AD significantly improved contouring accuracy compared to MD, with a mean DSC of 0.89 vs. 0.85 (p = 0.001), and reduced contouring time by an average of 53.5 % (193 s vs. 424 s, p < 0.001). Personalized DLAS model may increase the accuracy and efficiency of MD to optimize the workflow of MRgART for large BMs.