Click here to

Session: Multi-Disciplinary: Data Science/Radiomics [Return to Session]

Exploratory Unsupervised Structure-Learning Based Radiomics Approach for Brain Metastases Treatment Response Modeling of Stereotactic Radiosurgery

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


MO-IePD-TRACK 4-5 (Monday, 7/26/2021) 3:00 PM - 3:30 PM [Eastern Time (GMT-4)]

Purpose: Stereotactic radiosurgery (SRS) has been shown to reduce neurocognitive decline for patients with multiple (>4) brain metastases (BMs). Early, accurate and noninvasive treatment response diagnosis is essential for SRS management to assist subsequent therapy decision making. In this study, we develop a novel unsupervised structure learning-based radiomics approach for BMs treatment response modeling.

Methods: The study cohort includes 14 BMs SRS patients (307 BMs) with planning MRIs and dose distribution available. BMs segmentations are generated via our En-DeepMedic segmentation platform. For each BMs, 791 features are collected from MRI and dose distribution including clinical parameters, geometry features, first-order and texture features, using 7 types of ROI masks with different dilation margins. An unsupervised structure learning algorithm integrating both global and local information of the input is implemented. After optimization, a tree graph with skeleton and clusters on branches is learned to represent the intrinsic structure of the dataset in a reduced dimension, where similar subjects are clustered together and the main trajectory of the graph represents the inherent data structure. In addition, the algorithm can also perform clustering task based on input features.

Results: This unsupervised algorithm can discover a stable latent tree structure suggesting a linear progression trajectory starting from no size-change BMs towards size-changed BMs. The main skeleton diverges into three sub-branches in size-changed group, suggesting that possible responses subtypes can be further distinguished in that group. The algorithm can also cluster size-changed and no-change BMs into different subsets with F-measure score and Rand index as 0.70 and 0.65 respectively without any supervision. Specifically, no size-change group is clustered with only 6 misclassification.

Conclusion: This proposed unsupervised structure learning-based radiomics approach can reveal a tree structure with clusters associated with BMs treatment responses. This indicates that BMs treatment responses are separable only utilizing treatment images and dose.

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



    Stereotactic Radiosurgery, Modeling, Tumor Control


    TH- Response Assessment: Modeling: Machine Learning

    Contact Email