Pancreatic Mass Segmentation Using TransUNet Network

Fael Faray de Paiva, Alexandre de Carvalho Araujo, João Dallyson Sousa de Almeida, Anselmo C. de Paiva

2025

Abstract

Currently, one of the major challenges in computer vision applied to medical imaging is the automatic segmentation of organs and tumors. Pancreatic cancer, in particular, is extremely lethal, primarily due to the major difficulty in early detection, resulting in the disease being identified only in advanced stages. Recently, new technologies, such as deep learning, have been used to identify these tumors. This work uses the TransUNet network for the task, as convolutional neural networks (CNNs) are extremely effective at capturing features but present limitations in tasks that require greater context. On the other hand, transformer blocks are designed for sequence-to-sequence tasks and have a high capacity for processing large contexts; however, they lack spatial precision due to the lack of detail. TransUNet uses the Transformer as an encoder to enhance the capacity to process content globally, while convolutional neural networks are employed to minimize the loss of features during the process. Among the experiments presented herein, one used image pre-processing techniques and achieved an average Dice score of 42.60±1.97%. The second experiment, a crop was applied to the mass region, reaching an average Dice score of 79.67±2.31%.

Paper Citation

in Harvard Style

Faray de Paiva F., Araujo A., Sousa de Almeida J. and C. de Paiva A. (2025). Pancreatic Mass Segmentation Using TransUNet Network. In Proceedings of the 27th International Conference on Enterprise Information Systems - Volume 1: ICEIS; ISBN 978-989-758-749-8, SciTePress, pages 512-522. DOI: 10.5220/0013292200003929

in Bibtex Style

@conference{iceis25,
author={Fael Faray de Paiva and Alexandre Araujo and João Sousa de Almeida and Anselmo C. de Paiva},
title={Pancreatic Mass Segmentation Using TransUNet Network},
booktitle={Proceedings of the 27th International Conference on Enterprise Information Systems - Volume 1: ICEIS},
year={2025},
pages={512-522},
publisher={SciTePress},
organization={INSTICC},
doi={10.5220/0013292200003929},
isbn={978-989-758-749-8},
}

in EndNote Style

TY - CONF

JO - Proceedings of the 27th International Conference on Enterprise Information Systems - Volume 1: ICEIS
TI - Pancreatic Mass Segmentation Using TransUNet Network
SN - 978-989-758-749-8
AU - Faray de Paiva F.
AU - Araujo A.
AU - Sousa de Almeida J.
AU - C. de Paiva A.
PY - 2025
SP - 512
EP - 522
DO - 10.5220/0013292200003929
PB - SciTePress