A Smart Hybrid Framework for 3D Lung Nodule Segmentation Using Computed Tomography

Ahed Abugabah, Zayed University

Document Type

Article

Source of Publication

SN Computer Science

Publication Date

12-2-2025

Abstract

Lung cancer continues to be among the leading causes of cancer deaths worldwide, and so cases of lung nodules must be accurately and early detected. The latest developments of deep learning have enhanced the detection of nodules in 3D Computed Tomography (CT) data, but still, there are issues connected to computational complexity and performance effectiveness. In this paper, a new hybrid network, TransAtenNet, has been developed, which is TransAtenNet based on the concept of Swin Transformer combined with a U-Net structure that could help to solve these problems and improve the performance of lung nodule segmentation in three dimensions. The Swin Transformer is utilized to extract both global and local contexts, thus making the model appropriately sensitive to differentiate lung nodules under a variety of imaging conditions. In order to enhance model efficiency further, lightweight convolutional blocks taken from EfficientNet are implemented in the architecture to minimise computational requirements without affecting segmentation accuracy. We also consider the replacement of the standard attention mechanisms with squeeze-and-excitation (SE) blocks to decrease model complexity yet preserve the advantages of modified features learning. The model is tested on the LIDC-IDRI dataset, which is characterized by a highly variable number of nodules, variations like the nodules, slice thickness, and even inter-observer variability. Quantitatively, the proposed TransAtenNet achieved superior segmentation performance on the LIDC-IDRI dataset with a Dice Similarity Coefficient (DSC) of 0.913 and Intersection over Union (IoU) of 0.867, outperforming all baseline models, including Swin-UNet 3D, TransBTS, ViT-UNet, nnU-Net, and UNet3D. The smaller number of model parameters, better understanding of features, and computational efficiency make TransAtenNet a viable approach in the early detection and diagnosis of lung cancer, and in turn, the clinical usability of the model due to the smaller resource consumption and increased segmentation accuracy.

ISSN

2662-995X

Publisher

Springer Science and Business Media LLC

Volume

6

Issue

8

Disciplines

Computer Sciences | Medicine and Health Sciences

Keywords

3D lung nodule segmentation, Computed tomography (CT), Deep learning, Early lung cancer detection, Medical image analysis, Swin transformer

Scopus ID

105023571491

Indexed in Scopus

yes

Open Access

no

Link to Full Text

Share

COinS