14 MACHINE-LEARNING MODELS FOR PREDICTION OF RADIONECROSIS IN PATIENTS UNDERGOING INTRACRANIAL STEREOTACTIC RADIOSURGERY FOR BRAIN METASTASIS

14 种用于预测接受颅内立体定向放射外科治疗脑转移瘤患者放射性坏死的机器学习模型

阅读:1

Abstract

PURPOSE: To utilize machine learning and radiomic analysis to construct a model that can be used as a predictive tool for radiation necrosis (RN) after stereotatic radiotherapy (SRT) for brain metastasis (BM). MATERIALS AND METHODS: Patients who had SRT in Vancouver, BC, between the years 2019-2023 were included. There are 31 lesions in the RN cohort, as defined as having radiologic evidence of tumor growth post SRT and spontaneous regression with or without steroids. In the control arm, there are 30 lesions which did not develop RN. Patients with preceding surgical resection or those who passed away within 6 months of receiving SRS were excluded. For each lesion, dosimetric and geometric parameters such as size of the gross tumor volume (GTV), planning target volume (PTV), conformity index (CI), gradient index (GI), and homogeneity index (HI) were extracted. We then constructed a random forest model with 50 estimators, a maximum depth of 5 as a regularization parameter, and an 80-20 train-test split to predict whether the patient develops RN in the region of the lesion. In the future, we intend to use deep-learning model with MRI imaging and dose distribution to expand our model. RESULTS: The RN group had an average GTV that was statistically larger than the control group (3.65cc, 1.12cc, p=0.03), as well as a larger PTV (4.9cc, 1.7cc, p=0.02). The RN and control group received an average Dmax of 35Gy (p=0.26). The biologically equivalent dose (EQD2 ), with an alpha/beta ratio of 2, given to the RN and control group were 92Gy and 90Gy respectively (p=0.38).However, the volume of normal brain receiving >12 Gy (V12) in the RN group was significantly larger compared to the control group (27.8cc, 9.8cc, p=0.04). There was no statistical difference between the RN and control group with respect to quality metrics including CI (0.9, 1.0, p=0.27), GI (10.7, 7.4, p=0.59), and HI (1.3, 1.3, p=0.31). The random forest model achieved an accuracy of 83% in predicting RN outcomes in the test dataset. The top five most important features in descending order were the EQD2/2, EQD2/12, gradient index, volume enclosed in the 50% isodose surface, and maximum dose in tumor as a percentage of the prescription dose, with feature importance scores of 0.115, 0.097, 0.091, 0.088 and 0.072 respectively. CONCLUSIONS: While this pilot project is still in its early stages, the preliminary findings reveal a model that may accurately predict RN in patients who receive SRS.

特别声明

1、本页面内容包含部分的内容是基于公开信息的合理引用;引用内容仅为补充信息,不代表本站立场。

2、若认为本页面引用内容涉及侵权,请及时与本站联系,我们将第一时间处理。

3、其他媒体/个人如需使用本页面原创内容,需注明“来源:[生知库]”并获得授权;使用引用内容的,需自行联系原作者获得许可。

4、投稿及合作请联系:info@biocloudy.com。