The University of Auckland
Browse

Mishaim Malik: SnE-VNet: A Deep Learning Model with Squeeze and Excitation for Improved 3D Stroke Lesion Segmentation

Download (2.29 MB)
poster
posted on 2024-10-21, 22:38 authored by Mishaim MalikMishaim Malik, Benjamin Chong, Justin FernandezJustin Fernandez, Vickie ShimVickie Shim, Alan Wang

Stroke, a cerebrovascular disease, is the third leading cause of long-term disabilities in adults. Stroke rehabilitation plays a vital role in assisting stroke survivors to regain some function and in returning to their daily lives. One of the key aspects of stroke research is to predict clinical outcomes a few months after the stroke onset. Accurately predicting such outcomes can allow clinicians to prepare more effective rehabilitation plans. Using neuroimaging-based features to perform outcome predictions has shown promising results. However, a lesion mask must be generated to extract lesion-based metrics. Generally, manual segmentation is performed to generate the lesion mask; however, deep-learning models have also shown potential. In this study, we presented a deep learning-based stroke segmentation model. The presented model uses squeeze and excitation layers to capture spatial features within the input T1. The squeeze and excitation (SnE) block was implemented within the residual connection, ensuring a smooth flow of information within the network. The presented network was trained and tested using the publically available dataset ATLAS (v2.0). We concluded that adding the SnE block within the network resulted in faster convergence and improved results over baseline models, with a 0.6791 dice coefficient.

History

Publisher

University of Auckland

Usage metrics

    SGS Research Showcase

    Licence

    Exports

    RefWorks
    BibTeX
    Ref. manager
    Endnote
    DataCite
    NLM
    DC