מושגי ליבה
The BraTS-Reg challenge aimed to establish a public benchmark environment for deformable registration algorithms to estimate correspondences between pre-operative and follow-up MRI scans of patients with diffuse brain gliomas.
תקציר
The BraTS-Reg challenge was organized to establish a public benchmark environment for deformable registration algorithms to estimate correspondences between pre-operative and follow-up MRI scans of patients with diffuse brain gliomas. The challenge utilized a curated multi-institutional dataset of 259 diffuse glioma patients, with each patient having a pair of pre-operative baseline and follow-up MRI scans. The dataset was divided into training, validation, and testing cohorts. Clinical experts annotated ground truth landmark points on anatomically distinct locations in the baseline and follow-up scans.
The participating teams were required to submit containerized algorithms that could automatically register the baseline and follow-up scans. The algorithms were evaluated based on the Median Euclidean Error (MEE) between the warped landmark locations and the ground truth, Robustness (R) in terms of the proportion of landmarks with improved alignment, and the smoothness of the displacement field. A consolidated BraTS-Reg score was introduced to rank the methods.
The top-performing methods shared several commonalities, including pre-alignment, deep neural networks, inverse consistency analysis, and test-time instance optimization. The best method achieved MEE at or below the inter-rater variability for approximately 60% of the evaluated landmarks, highlighting the potential for further accuracy and robustness improvements. The challenge data and online evaluation tools remain accessible to serve as an active resource for research.
סטטיסטיקה
"The time-window between all pairs of baseline and follow-up MRI scans was in the range of 27 days – 48 months."
"The total number of landmarks (χ+ψ) varied for each case between 6 and 50."
ציטוטים
"Registration is a fundamental problem in medical image analysis (Sotiras et al., 2013; Ou et al., 2014) that aims to find spatial correspondences between two images and align them for various downstream applications."
"Accurate longitudinal image registration between pre-operative and follow-up scans is particularly crucial for patients with brain tumors. Such registration can aid in analyzing the characteristics of healthy tissue, potentially identifying tumor recurrence (Han et al., 2020)."