A Self-Attention Network With Position-Prior Guided Feature Mapping And Clustering For 3D Pancreas Segnentation

• Published in: 2024 IEEE International Symposium on Biomedical Imaging (ISBI) pp. 1 - 5
• Main Authors: Fan, Kaiyang, Zhang, Min, Xia, Shouhai, Du, Junli, Yang, Pan, Zhang, Xiao, Feng, Jun
• Format: Conference Proceeding
• Language: English
• Published: IEEE 27.05.2024
• Subjects: Accuracy; Biological systems; Computed tomography; Decoding; Deformation; feature clustering; Feature mapping; Image segmentation; pancreas segmentation; position-prior guided; Three-dimensional displays
• ISSN: 1945-8452
• DOI: 10.1109/ISBI56570.2024.10635285

Accurate segmentation of small organs, such as the pancreas, gallbladder and the adrenal gland from CT scans is still a challenge task due to the category imbalance, large deformation, and blurred border in the medical images of small organs. This paper proposes a new self-attention network with position-prior guided feature mapping and clustering (PGFMC-Net) for 3D pancreas segmentation in abdominal CT scans. Particularly, a position-prior guided feature mapping (PGFM) module is incorporated into the encoder, reducing the effects of category imbalances. Auxiliary location path (ALP) and deformable convolution are introduced in the PGFM module to integrate the position information of the feature map. Meanwhile, a position-prior guided feature clustering (PGFC) module has also been introduced in the decoder to solve the difficulties of large deformation and blurred border in pancreas segmentation. We evaluate the proposed method on three public pancreas segmentation datasets, and the experimental results show that our method achieves the best results on all three datasets.