Date of Award

January 2021

Document Type

Thesis

Degree Name

Medical Doctor (MD)

Department

Medicine

First Advisor

Sanjay Aneja

Abstract

It is a challenge to model survival for patients with brain metastases given their clinical heterogeneity. Quantitative imaging biomarkers, including radiomic features, have shown promise for modeling cancer outcomes. In light of the ever-increasing amount of medical data, deep learning - a branch of machine learning - is well-suited for modeling high-dimensional non-linear relationships. We hypothesized that a deep learning survival model incorporating quantitative imaging biomarkers would be more effective than traditional models of survival in patients with brain metastases.

We analyzed 831 patients with 3596 total brain metastases treated with primary stereotactic radiosurgery at our institution between 2000-2018. The primary outcome of interest was overall survival following treatment. Clinical variables included age, Karnofsky performance status, presence of extracranial metastases, and number of brain metastases. 851 3D radiomic features were extracted from pre-treatment T1 post-contrast MRI images of each brain metastasis and aggregated per patient. Traditional Cox proportional hazards models and deep learning survival models were developed with clinical variables alone, relevant imaging features alone, and a combination of clinical and imaging features. Performance of each model was assessed with Harrell’s Concordance-Index.

Median overall survival was 12 months. Median age was 63 years. Median follow-up was 36 months. The most common primary sites were NSCLC (38.5%), melanoma (18.9%), breast (14.9%), SCLC (7.2%), renal (5.3%), and GI (4.7%). The traditional model trained with clinical variables alone, radiomic features alone, and both clinical and radiomic features, had C-Indices of 0.644, 0.559, and 0.670, respectively. The deep learning survival model incorporating both clinical variables and radiomic features performed best overall with a C-Index of 0.887 compared to deep learning survival models utilizing clinical variables alone and radiomic features alone (C-Indices 0.653 and 0.788).

A deep learning model incorporating clinical data with quantitative radiomic imaging features performed better than a traditional linear model using clinical predictors alone at modeling survival in patients with multiple brain metastases. This represents a promising method to personalize prognostication with implications for guiding goals of care conversations and increasing clinical trial eligibility and risk stratification.

Comments

This thesis is restricted to Yale network users only. This thesis is permanently embargoed from public release.

Share

COinS