Intranet: infrared-based transformers for 2D medical image segmentation

Lin, Hangzheng

Permalink

https://hdl.handle.net/2142/120132 Copy

Description

Title

Intranet: infrared-based transformers for 2D medical image segmentation

Author(s)

Lin, Hangzheng

Issue Date

2023-05-01

Director of Research (if dissertation) or Advisor (if thesis)

Kindratenko, Volodymyr

Department of Study

Electrical & Computer Eng

Discipline

Electrical & Computer Engr

Degree Granting Institution

University of Illinois at Urbana-Champaign

Degree Name

M.S.

Degree Level

Thesis

Keyword(s)

• Machine Learning
• Biomedical
• Infrared Imaging
• Segmentation
• Transformer
• Deep Learning

Abstract

Infrared (IR) spectroscopic imaging is widely employed in medical imaging applications due to its ability to capture both chemical and spatial information of biological tissues. In recent years, convolutional neural networks (CNNs), including the well-known U-Net model, have demonstrated impressive performance in biomedical image segmentation. However, the inherent locality of convolution limits the effectiveness of these models for encoding IR data, resulting in suboptimal performance in for some applications. In this work we propose an infrared-based transformer network named INTRANET for IR image segmentation. This novel model leverages the strength of the transformer encoders to segment infrared colon images effectively. Incorporating the skip-connection and transformer encoders, INTRANET overcomes the issue of pure convolution models, such as the difficulty of capturing long-range dependencies. We train several encoder-decoder models on a colon dataset of IR images to evaluate the existing convolution models and our proposed method. Our model achieves an AUC score of 0.9872, using 17 spectral bands for the segmentation task. Experimental results demonstrate that INTRANET significantly improves over the pure convolution models, especially when the input IR band number is limited.

Graduation Semester

2023-05

Type of Resource

Thesis

Copyright and License Information

Copyright 2023 Hangzheng Lin

Owning Collections