Lauren M Kim MD, Presenter: Nothing to Disclose

Minmin Chen, Abstract Co-Author: Nothing to Disclose

Kilian Weinberger, Abstract Co-Author: Nothing to Disclose

Paul K. Commean, Abstract Co-Author: Nothing to Disclose

Parinaz Massoumzadeh PhD, Abstract Co-Author: Nothing to Disclose

Tammie Smith Benzinger MD, PhD, Abstract Co-Author: Research grant, Eli Lilly and Company

Fred William Prior PhD, Abstract Co-Author: Research collaboration, Carestream Health, Inc Research partner, Merge Healthcare Research collabortation, Siemens AG

Sarah C. Jost MD, Abstract Co-Author: Nothing to Disclose

Daniel Marcus PhD, Abstract Co-Author: Owner, Radiologics, Inc

Glioblastoma multiforme (GBM) is the most common primary malignant neoplasm of the adult brain. The permeative nature of GBM makes tumor margin delineation challenging, at times limiting the efficacy of surgical resection. An improved method of assessing the geographic extent of GBM is therefore warranted. The authors hypothesize that a machine learning algorithm (MLA) can be trained to accurately segment a GBM patient’s MRI examination once it has been trained using a radiologist’s segmentations.

Data were chosen from the COmprehensive Neuro-oncology Data Repository (CONDR) at Washington University in St. Louis and Swedish Neuroscience Institute. MRI sequences (pre- and post-contrast T1WI, FLAIR, DWI, ADC, FA, SWI, rCBV, rCBF, MTT, TTP, and MPRAGE) were resized to cubic voxels and coregistered. A radiologist segmented the lesion and predicted the presence of malignant tissue on a voxel-by-voxel basis, ranging from high probability to very low probability, using pre- and post-contrast T1WI, ADC, FLAIR, SWI, and rCBV sequences. In a leave one out study design, a random forests MLA was trained using the radiologist’s classifications as truth.

When the trained classifier was applied on a voxel-by-voxel basis to the 12 MRI sequences of a test subject’s brain, an accurate segmentation prediction resulted. ROC analysis was used to evaluate the accuracy of the MLA to correctly segment the lesion from the brain in test subjects compared to the predictions of a radiologist. The area under the curve (AUC) was used to quantify the high precision of the prediction, and AUC values of 0.95 were achieved.

Given the challenges implicit in the invasive pathophysiology of GBM and poor clinical outcomes of these patients under existing standard of care paradigms, a new approach in the radiologic assessment of GBM is warranted. The MLA is able to accurately segment a GBM lesion, using a radiologist as a truth standard. Future directions include validating the diagnostic accuracy of the MLA with histopathologic analysis of biopsy specimens of the patients included in this study. With such validation, the MLA may serve as a powerful clinical tool in the evaluation and surgical treatment of GBM.

The MLA provides a way of combining data from multiple MRI sequences to produce a computer-generated composite prediction which may aid in the surgical biopsy and/or resection of GBM.

Kim, L, Chen, M, Weinberger, K, Commean, P, Massoumzadeh, P, Benzinger, T, Prior, F, Jost, S, Marcus, D, Machine Learning Algorithm for the Magnetic Resonance Characterization of Glioblastoma Mulitforme.  Radiological Society of North America 2011 Scientific Assembly and Annual Meeting, November 26 - December 2, 2011 ,Chicago IL. http://archive.rsna.org/2011/11015297.html