Automatic Brain Extraction Using Deep Learning

AbstractBackgroundBrain extraction from cranial MR images is an important precursor to many medical‐imaging algorithms prominent in Alzheimer’s disease research, including image registration, atlas‐based segmentation and longitudinal analyses. Moreover, skull stripping can be an important step in deidentification by removing facial features from the image. One widely used method is PiNCRAM, which provides reliable results, however, is computationally intensive and time consuming. This is a bottleneck for scalable clinical trial data processing and potential clinical applications. The emergence of deep‐learning segmentation tools (convolutional neural networks – CNNs) provide a framework for drastically reducing the processing time and potentially increasing the reproducibility. In this work, we develop a CNN‐based skull‐stripping method that mimics PiNCRAM, but runs in a fraction of the time.MethodA cohort of 2300 T1 weighted scans from healthy participants with a slightly increased risk for Alzheimer’s Disease, drawn from the European Prediction of Alzheimer’s Dementia (EPAD) consortium was used to train a CNN. The examples fed to the model are generated from PiNCRAM masks. The network architecture is a 3D U‐NET consisting of around 200,000 parameters and was trained using the ADAM optimiser with default settings (as implemented in Tensorflow 2).We used 50 T1 weighted scans from the Open Access Series of Imaging Studies (OASIS) test‐retest cohort as a validation set.ResultVisually, the CNN is seen to segment the brain well (Fig 1). The mean dice overlap score on the validation set is 0.9811 (std. 0.002791). The median CNN test‐retest volume change percentage (expressed between 0 and 1) was ‐0.000854 (std. 0.006703), whilst PiNCRAM was 0.00883 (std. 0.0115; Fig. 2, Table 1). From the superior reproducibility, one could infer that dice overlap less than unity does not imply CNN error (Fig 1.)Our CNN pipeline takes approximately 1.1% of the time taken by PiNCRAM. Each brain takes approximately 20 seconds (when in batch mode) versus 30 minutes for PiNCRAM.ConclusionThe PiNCRAM‐CNN provides a robust and highly accurate brain extraction alternative to PiNCRAM. Not only is it two orders of magnitude faster to run, but it also displays an order of magnitude better reproducibility.