Adaptive Kernel-Based Serial and Parallel U-Net Architectures for the Segmentation of Lung Nodules

R. Janefer Beula () and A. Boyed Wesley ()
Additional contact information
R. Janefer Beula: Marthandam Affiliated to Manonmaniam Sundaranar University
A. Boyed Wesley: Marthandam Affiliated to Manonmaniam Sundaranar University

Computational Statistics, 2025, vol. 40, issue 9, No 27, 5617-5647

Abstract: Abstract This paper proposes a 2D-lung nodule segmentation approach that uses adaptive kernel-based serial (ADS) and parallel (ADP) U-Net architectures. This approach initially pre-processes the CT scan images of the lungs and segments the lung nodule candidates. The pre-processed images are further subdivided into non-overlapping patches, and the patches are classified as essential and non-essential patches based on the lung nodule candidates. Essential patches are the patch that has part or full nodule candidates, while the non-essential patches do not contain a nodule candidate. The essential patches are combined and fed to the ADS-U-Net or ADP-U-Net architectures to detect the actual lung nodules. The proposed Serial or parallel U-Net architectures have two sections of the U-Net architecture. The ADS-U-Net uses one U-Net architecture with convolutional kernels in series with the U-Net architecture with adaptive kernels. Instead, the ADP-U-Net uses the two U-Net architectures in parallel. The difference between the adaptive and convolutional kernels in the U-Net architectures is that the coefficients of the kernel remain constant in the convolutional kernel, which is only updated during backpropagation, while the coefficients of the adaptive kernel get updated throughout the convolution striding process. The proposed two lung nodule segmentation approaches, ADS-U-Net and ADP-U-Net, are evaluated using the LIDC and the Cancer Imaging datasets, which contain CT scans from 1018 and 355 subjects, respectively. The measures such as accuracy, sensitivity, dice similarity coefficient (DSC), positive predictive value (PPV), Hausdorff distance (HD), and probability rand index (PRI) are considered for evaluation. The proposed ADS-U-Net architecture results in an accuracy, Sensitivity, DSC, and PPV of 92.18%, $$90.26\%$$ , 9 $$0.24\%$$ , and $$81.48\%$$ respectively when evaluated using the LIDC dataset. With the same LIDC dataset, the ADP-U-Net architecture yields an accuracy, Sensitivity, DSC, and PPV of 92.63%, $$90.59\%$$ , 9 $$0.61\%$$ , and $$81.83\%$$ respectively. In the case of the Cancer imaging dataset, the ADS-U-Net and ADP-U-Net architecture results in a segmentation accuracy of 92.96% and 93.74%, respectively. When comparing the two architectures, the ADP-U-Net results in better performance during cross-dataset evaluation than the ADS-U-Net.

Keywords: Convolutional kernel; U-Net; Lung nodule candidates; Lung parenchyma; CT images (search for similar items in EconPapers)
Date: 2025
References: Add references at CitEc
Citations:

Downloads: (external link)
http://link.springer.com/10.1007/s00180-025-01668-5 Abstract (text/html)
Access to the full text of the articles in this series is restricted.

Related works:
This item may be available elsewhere in EconPapers: Search for items with the same title.

Export reference: BibTeX RIS (EndNote, ProCite, RefMan) HTML/Text

Persistent link: https://EconPapers.repec.org/RePEc:spr:compst:v:40:y:2025:i:9:d:10.1007_s00180-025-01668-5

Ordering information: This journal article can be ordered from
http://www.springer.com/statistics/journal/180/PS2

DOI: 10.1007/s00180-025-01668-5

Access Statistics for this article

Computational Statistics is currently edited by Wataru Sakamoto, Ricardo Cao and Jürgen Symanzik

More articles in Computational Statistics from Springer
Bibliographic data for series maintained by Sonal Shukla () and Springer Nature Abstracting and Indexing ().