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

Congratulations to Drs. David E Vaillancourt and Steven Trent DeKosky on the publication of “Automated Hippocampus Segmentation and Volume Estimation Using a Transformer-based Deep Learning Architecture,” which appears in pre-print  at Research Square.



Hippocampus segmentation in brain MRI is a critical task for diagnosis, prognosis, and treatmentplanning of several neurological disorders. However, automated hippocampus segmentation methodshave some limitations. More precisely, hippocampus is hard to visualize through MRI due to the lowcontrast of the surrounding tissue, also it is a relatively small region with highly variable shape. In thisstudy, we propose a two-stage architecture to rst locate the hippocampus and then segment it. Wecombine a transformer design with CNN based architecture and a customized loss function to segmentthe hippocampus via an end-to-end pipeline. In the encoding path, the image is passed through a CNNmodel to generate a feature map. This feature map is then divided into small patches which are passedto a transformer for extracting global contexts. The encoder used here is identical to that of the VisionTransformer image classication model. In the decoding path, the transformer outputs are combined withtheir corresponding feature maps to enable a precise segmentation of the hippocampus. The proposedarchitecture 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 thissegmentation task. The mean volume difference between generated mask and ground truth is 5% with astandard deviation of 3%. Deploying the proposed method over our in-house data, consisting of 326MRIs, showed a mean volume difference of 4.4 % with a standard deviation of 3.24%