Session: 02-09-01: Design for Healthcare Products and Processes

Paper Number: 150328

150328 - Segmentation of Gastrointestinal Cancer Using Trans-Unet and Mri Images 

Gastro-Intestinal (GI) tract cancers affect an estimated 5 million individuals globally each year,
accounting for over a quarter of all cancer cases. These cancers typically originate from genetic
mutations in the cells lining the digestive tract, leading to abnormal cell growth and eventually
cancer. The primary treatment strategy for GI tract cancers involves surgical resection of the
primary tumor and regional lymph nodes, along with chemotherapy and radiotherapy.
Recent advancements in deep learning, particularly in segmentation U-Nets, have
demonstrated significant potential in automating the segmentation process. This automation
reduces manual labor and enables more patients to receive appropriate treatment. Deep neural
network-based methods are increasingly employed for the automated diagnosis of medical
illnesses, outperforming traditional algorithms in accuracy due to their data-driven nature.
These methods learn characteristics from data using general-purpose learning techniques
rather than relying on manually crafted features.
In medical imaging, segmentation involves dividing an image into multiple segments or
regions, each representing different parts of the body or anatomical structures. This technique
is essential for identifying and isolating specific areas of interest, such as tumors or organs.
This research focuses on enhancing the segmentation accuracy of MRI images for GI tract
cancers, which is critical for effective radiation therapy planning. Employing advanced deep
learning techniques, particularly Transformer-based architectures integrated with U-Net
models, this study aims to develop a segmentation model that surpasses the current state-ofthe-
art in accuracy. The proposed model leverages the inherent strengths of Transformers to
manage the complex spatial relationships in MRI scans, facilitating more precise targeting of
radiation doses and potentially improving patient outcomes.
Extensive experiments have demonstrated that Transformer U-Nets outperform other methods
in various medical applications, highlighting the effectiveness of Transformers and sophisticated
attention components in medical segmentation. In our work, we achieved significant
improvements in accuracy measurements, such as DICE and Intersection over Union (IoU)
scores, surpassing the current state-of-the-art methods like BiFTransNet. By optimizing fusion
operations and hyperparameters, our model provides substantial improvements over existing
techniques. Moreover, we have developed a working system that not only accurately segments
MRI images but also performs this operation in a timely manner, ensuring acceptable inference
times.
As the prevalence of GI tract cancers continues to rise, integrating sophisticated techniques like
these into clinical practice becomes increasingly vital. Our findings contribute to the growing
body of evidence supporting the use of deep learning in medical imaging and highlight the need
for ongoing research and development in this field. These advancements could lead to better
diagnostic capabilities and treatment planning for GI tract cancers, ultimately improving patient
outcomes.
The research underscores the potential of advanced deep learning methods, particularly
Transformer-based U-Nets, in enhancing the accuracy and efficiency of medical image
segmentation. By leveraging the strengths of these advanced architectures, our model can
manage complex spatial relationships in MRI scans, leading to more precise targeting of
radiation doses and potentially improved patient outcomes. As the incidence of GI tract cancers
continues to rise, integrating such sophisticated techniques into clinical practice is increasingly
vital. Our work contributes to the growing body of evidence supporting the use of deep learning
in medical imaging and underscores the need for ongoing research and development in this
critical area.

Presenting Author: Miri Weiss Cohen Braude College of Engineering

Presenting Author Biography: Miri Weiss Cohen is in the Head of Department of Software Engineering at the Braude College
of Engineering. Her research interests lie in the area Machine Learning (AI) methodologies in
use of optimization for engineering problems. In recent years, she is focusing on two major
topics: first, prediction and optimization of green energy (solar and wind) models using big
data time series. Second, using CT and MRI scans for classification and segmentation of cancer
, reconstructing 3D volumes Deep learning. She has collaborated actively with researchers in
several other disciplines relating to Machine Learning in service of rehabilitation and design of
human computer interfaces

Authors:

Miri Weiss Cohen Braude College of Engineering