Using Prognostic Indices for Brain Metastases in the Setting of Delayed Stereotactic Radiosurgery (SRS)

Abstract

Purpose/Objective(s): Conventional MRI is no longer sufficient to accurately identify tumor presence considering the widely documented infiltrative nature of gliomas. Diffusion Weighted Imaging (DWI) has lent further interpretation to available imaging, but data on the optimal combination of DWI sequences to be employed remains elusive. Our intent was to use a convolutional neural network approach to machine learning employing DWI sequences (apparent diffusion coefficient (ADC), relative cerebral blood volume (rCBV)) to identify an imaging biomarker for GBM. Materials/Methods: Ten histologically documented GBM cases with available detailed operative reports and gross tumor present on DWI prior to the administration of radiation therapy were selected. T1 post-gad images were used to manually delineate tumor which were then coregistered with DWI series. These 10 manually delineated tumors were used to train a convolutional neural network classifier (CNN). In testing, the trained CNN is employed to assign each pixel in the image a probability of belonging to tumor. Receiver Operating Characteristic (ROC) analysis was performed on the probability map to determine optimum thresholds for tumor grading and to obtain the sensitivity, specificity, and positive and negative predictive values for identifying high-grade gliomas. Results: Backtesting of the 10 GBM cases used for machine training achieved almost 100% probability concordance with the T1-gad manually delineated tumors. The trained CNN was then tested on 5 GBM patient datasets with T1-gad, ADC, rCBV images. A sensitivity of 75% and a specificity of 80% with positive and negative predictive values of 79% and 76% respectively were achieved. Optimum threshold for tumor was 0.47. Conclusion: The convolutional neural network approach to DWI analysis may be very useful in identifying a high grade glioma imaging biomarker. Further training with patient data will further improve the accuracy of this approach, enabling its use for recurrence pattern analysis in setting of radiation therapy and systemic treatment, with possible future applicability in tumor grading as well as radiation treatment field design.Purpose/Objective(s): SRS is an important tool in treating brain metastases and has made modern treatment options more sophisticated. Some patients may benefit from radiosurgery at the initial diagnosis of brain metastasis while others may benefit from delayed SRS after the use of whole brain radiation (WBRT) or surgery. Prognostic scoring systems help select patients who will benefit from SRS by estimating their survival.However, because these indices were validated to estimate survival only at the time of initial diagnosis of brain metastases, it is unclear whether or not they remain prognostic when used months after the discovery of brain metastases. In this study, we assess these prognostic indices in patients evaluated for SRS delayed from their initial diagnosis. Materials/Methods: We reviewed all patients treated at our institution with SRS for brain metastases ...