Abstract

In computed tomography (CT) imaging, the Hounsfield Unit (HU) scale quantifies radiodensity, but its nonlinear nature across organs and lesions complicates machine learning analysis. This paper introduces an automated method for adaptive HU scale windowing in deep learning-based CT liver segmentation. We propose a new neural network layer that optimizes HU scale window parameters during training. Experiments on the Liver Tumor Segmentation Benchmark show that the learned window parameters often converge to a range encompassing clinically used windows but wider, suggesting that adjacent data may contain useful information for machine learning. This layer may enhance model efficiency with just 2 additional parameters.

Recommended Citation

Zakrzewski, M., Kwiatkowski, D. & Cychnerski, J. (2024). Adaptive Hounsfield Scale Windowing in Computed Tomography Liver Segmentation. In B. Marcinkowski, A. Przybylek, A. Jarzębowicz, N. Iivari, E. Insfran, M. Lang, H. Linger, & C. Schneider (Eds.), Harnessing Opportunities: Reshaping ISD in the post-COVID-19 and Generative AI Era (ISD2024 Proceedings). Gdańsk, Poland: University of Gdańsk. ISBN: 978-83-972632-0-8. https://doi.org/10.62036/ISD.2024.8

Paper Type

Poster

DOI

10.62036/ISD.2024.8

Download

Share

COinS
 

Adaptive Hounsfield Scale Windowing in Computed Tomography Liver Segmentation

In computed tomography (CT) imaging, the Hounsfield Unit (HU) scale quantifies radiodensity, but its nonlinear nature across organs and lesions complicates machine learning analysis. This paper introduces an automated method for adaptive HU scale windowing in deep learning-based CT liver segmentation. We propose a new neural network layer that optimizes HU scale window parameters during training. Experiments on the Liver Tumor Segmentation Benchmark show that the learned window parameters often converge to a range encompassing clinically used windows but wider, suggesting that adjacent data may contain useful information for machine learning. This layer may enhance model efficiency with just 2 additional parameters.