Stable Atlas-based Mapped Prior (STAMP) machine-learning segmentation for multicenter large-scale MRI data

Machine learning (ML)-based segmentation methods are a common technique in the medical image processing field. In spite of numerous research groups that have investigated ML-based segmentation frameworks, there remains unanswered aspects of performance variability for the choice of two key components: ML algorithm and intensity normalization. This investigation reveals that the choice of those elements plays a major part in determining segmentation accuracy and generalizability. The approach we have used in this study aims to evaluate relative benefits of the two elements within a subcortical MRI segmentation framework. Experiments were conducted to contrast eight machine-learning algorithm configurations and 11 normalization strategies for our brain MR segmentation framework. For the intensity normalization, a Stable Atlas-based Mapped Prior (STAMP) was utilized to take better account of contrast along boundaries of structures. Comparing eight machine learning algorithms on down-sampled segmentation MR data, it was obvious that a significant improvement was obtained using ensemble-based ML algorithms (i.e., random forest) or ANN algorithms. Further investigation between these two algorithms also revealed that the random forest results provided exceptionally good agreement with manual delineations by experts. Additional experiments showed that the effect of STAMP-based intensity normalization also improved the robustness of segmentation for multicenter data sets. The constructed framework obtained good multicenter reliability and was successfully applied on a large multicenter MR data set (n > 3000). Less than 10% of automated segmentations were recommended for minimal expert intervention. These results demonstrate the feasibility of using the ML-based segmentation tools for processing large amount of multicenter MR images. We demonstrated dramatically different result profiles in segmentation accuracy according to the choice of ML algorithm and intensity normalization chosen.