Click here to


Are you sure ?

Yes, do it No, cancel

A Deep Learning Based Segmentation and Evaluation Framework for Brain Metastases Follow-Up After Stereotactic Radiosurgery

Z Yang1*, L Wang2, Y Liu3, M Chen1, E Zhang1, R Timmerman1, T Dan1, Z Wardak1, W Lu1, X Gu1, (1) UT Southwestern Medical Center, Dallas, TX (2) University Of Texas At Arlington, TX (3) Sichuan University, Chengdu, CN


(Monday, 7/13/2020) 4:30 PM - 5:30 PM [Eastern Time (GMT-4)]

Room: Track 2

Purpose: As a stand of care for brain metastases (BMs), stereotactic radiosurgery (SRS) enables the treatment of multiple (>4) BMs (mBMs) patients with excellent tumor control. In this study, we propose a deep learning driven systematic BMs follow-up evaluation framework to ensure the treatment quality and post-treatment quality of life (QoL).

Methods: The framework consists of a segmentation session and an evaluation session. A) In the segmentation session, BMs are auto-segmented in follow-up MRIs using our En-DeepMedic CNN-based segmentation platform. Deformable image registration is utilized to help excluding false-positives and identifying recurrences by mapping segmentations from planning MRI. BMs volumes are calculated for every follow-up. B) In the evaluation session, BMs post-treatment progression pattern is analyzed. First, polynomial curve fitting and dynamic time warping is conducted to obtain the volumes of each BMs changing over time. K-means clustering is then performed to cluster BMs into different progression categories and discover representative patterns, which can be future used for treatment outcome prediction. Thirteen mBMs (vary from 7-69) SRS patients with follow-up MRIs are collected to test this framework.

Results: For the segmentation session, our approach can detect mBMs on follow-up MRIs with averaged sensitivity of 0.97. Segmentation accuracy is evaluated by averaged center of mass shift as 1.41±0.32mm, Hausdorff distance as 2.75±0.57mm, the mean of surface-to-surface distance (SSD) as 1.01±0.28mm and the standard deviation of SSD as 0.80±0.19mm. For the evaluation session, our approach can differentiate BMs into different post-treatment progression trend categories, and reveal the representative patterns in each categories to aid treatment outcome prediction.

Conclusion: The developed deep learning based systematic follow-up framework can accurately identify mBMs and conduct tumor progression/regression analysis based on segmented volumes. It can assist BMs’ follow-up clinical workflow and aid BM SRS treatment outcome prediction, and is a promising tool for BM SRS management.

Funding Support, Disclosures, and Conflict of Interest: This work is supported by NIH R01 CA235723 and the seed grant from Department of Radiation Oncology at University of Texas Southwestern Medical Center.


Segmentation, Gamma Knife, Tumor Control


TH- Response Assessment: Modeling: Machine Learning

Contact Email