Towards a generalized deep learning framework for production scale segmentation of brain structures

Abstract

AbstractBackgroundWhole brain segmentation from magnetic resonance imaging (MRI) remains a vital step in analysis workflows. We developed a deep learning framework that dramatically reduces the computational and human resources required for removal of non‐brain tissues. Segmentation of brain structures is currently resource intensive and remains a bottleneck in the analysis pipeline. Here we report on the capabilities of our previously developed deep learning framework to segment three specific brain structures of varying difficulty.MethodWe trained and tested a deep learning framework on three brain structures of varying difficulty: the Hippocampus, Cortical Grey Matter, and Lateral Ventricles. Training set size differed between segmentation task based on ground truth availability: Hippocampus (8,780 train, 878 test), Cortical Grey Matter (8,900 train, 890 test), Lateral Ventricle (8,700 train, 870 test). Data sets were derived from the following studies: ADNI, ADC, Khandle, Sol‐Inca and 90plus. At the core of our framework is a tunable multi‐stage convolutional neural network (CNN). The configuration used for these experiments is the same as our top performing configuration used for whole brain segmentation, a 5 stage 13 layer encoder followed by a 6 layer fusion decoder. All experiments were performed on an NVIDIA DGX Station with four Tesla V100 16GB GPUs. For both training and testing we used ground truth masks generated by atlas‐based segmentation techniques followed by human quality control. We used the F1 score, or dice similarity coefficient, between each prediction and ground truth mask to evaluate the quality.ResultSample predictions and their corresponding ground truth masks are shown in Figures 1‐3. F1 score distributions are shown in Figure 4 indicating performance in diverse imaging cohorts. General performance is high, but varies amongst ROIs. Outliers are most likely explained by low quality data.ConclusionOur preliminary results suggest our current deep learning framework can generalize relatively well without modification to its configuration, although performance for each ROI varies. Further modifications to the framework are needed to achieve production scale quality as well as increased quality control of ground truth masks to reduce the number of outliers and improve general performance.