Browse by author
Lookup NU author(s): Dr Shidong WangORCiD
This work is licensed under a Creative Commons Attribution 4.0 International License (CC BY 4.0).
© 2024 Elsevier B.V.Prototypical networks have emerged as the dominant method for Few-shot Medical image Segmentation (FSMIS). Despite their success, the commonly used Masked Average Pooling (MAP) approach in prototypical networks computes the mean of the masks, resulting in imprecise and inadequate prototypes that fail to capture the subtle nuances and variations in the data. To address this issue, we propose a simple yet effective module called De-biasing Masked Average Pooling (DMAP) to generate more accurate prototypes from filtered foreground support features. Specifically, our approach introduces a Learnable Threshold Generation (LTG) module that adaptively learns a threshold based on the extracted features from both support and query images, and then choose partial foreground pixels that have larger similarity than the threshold to generate prototypes. Our proposed method is evaluated on three popular medical image segmentation datasets, and the results demonstrate the superiority of our approach over some state-of-the-art methods. Code is available at https://github.com/YazhouZhu19/DMAP.
Author(s): Zhu Y, Cheng Z, Wang S, Zhang H
Publication type: Article
Publication status: Published
Journal: Pattern Recognition Letters
Year: 2024
Volume: 183
Pages: 71-77
Print publication date: 01/07/2024
Online publication date: 09/05/2024
Acceptance date: 05/05/2024
Date deposited: 22/07/2024
ISSN (print): 0167-8655
ISSN (electronic): 1872-7344
Publisher: Elsevier B.V.
URL: https://doi.org/10.1016/j.patrec.2024.05.003
DOI: 10.1016/j.patrec.2024.05.003
ePrints DOI: 10.57711/z2qg-gk21
Data Access Statement: The employed datasets are publicly available.
Altmetrics provided by Altmetric