IMG-29. IMPLICATIONS OF IMAGE HARMONIZATION FOR BRAIN TUMOR CLINICAL TRIALS

Clinical trials for rare pediatric conditions like optic pathway gliomas associated with NF1 (NF1-OPGs) require numerous sites that utilize different MRI scanners and protocols. These MRIs must be harmonized for reliable comparison using machine learning tools. We propose a deep-learning based MRI harmonization that enables reproducible data analysis across multiple sites and MRI platforms.

One-hundred eighty clinical T1-weighted-MRIs of pediatric NF1-OPG subjects were acquired from three sites [Children’s National Hospital (site A, N=60), Children’s Hospital Colorado (site B, N=60), and Children’s Hospital of Philadelphia (site C, N=60)] using different MR scanners (GE, Phillips, and Siemens, respectively). MRIs from A and B were used to train a branched neural network and C were used for testing. Using unsupervised learning, the neural network identified differences in imaging protocols and harmonized data to a chosen protocol, e.g., that of site A, while preserving the patient anatomy. The harmonization was evaluated using Wasserstein distance (nWD), which measured similarity between two gray scale appearances of MRI intensity histograms. Anatomy preservation metric (AP) measured relative volume changes in gray matter (GM) segmented using Freesurfer, before and after harmonization. For both metrics, 100% means the best possible harmonization was achieved.

Before harmonization, nWD between grayscale appearance of MRIs from sites A and B was 27.3±3.1%. After harmonization, the similarity became 94.2±2.0% with a GM volume error of 3.11±0.56ml which corresponded to AP of 93.1±4.7%. For the test data, nWD similarity post harmonization improved from 38.2±3.9% to 93.4±2.2% with GM volume error of 4.12±0.47ml which corresponded to AP of 90.3±3.5%.

Our approach enables MRI harmonization while preserving anatomy for precise analysis using machine learning tools. It scales to new sites to support multisite clinical trials that utilize imaging outcomes for brain tumors and other rare diseases.