Automated Hippocampus Segmentation and Volume Estimation Using a Transformer-based Deep Learning Architecture

Abstract Hippocampus segmentation in brain MRI is a critical task for diagnosis, prognosis, and treatment planning of several neurological disorders. However, automated hippocampus segmentation methods have some limitations. More precisely, hippocampus is hard to visualize through MRI due to the low contrast of the surrounding tissue, also it is a relatively small region with highly variable shape. In this study, we propose a two-stage architecture to first locate the hippocampus and then segment it. We combine a transformer design with CNN based architecture and a customized loss function to segment the hippocampus via an end-to-end pipeline. In the encoding path, the image is passed through a CNN model to generate a feature map. This feature map is then divided into small patches which are passed to a transformer for extracting global contexts. The encoder used here is identical to that of the Vision Transformer image classification model. In the decoding path, the transformer outputs are combined with their corresponding feature maps to enable a precise segmentation of the hippocampus. The proposed architecture was trained and tested on a dataset containing 195 brain MRI from the Decathlon Challenge. The proposed network achieved a Dice value of 0.90±0.200, and 89% mean Jaccard value in this segmentation task. The mean volume difference between generated mask and ground truth is 5% with a standard deviation of 3%. Deploying the proposed method over our in-house data, consisting of 326 MRIs, showed a mean volume difference of 4.4 % with a standard deviation of 3.24%.