Related Experiment Video
Updated: Dec 27, 2025
10:26
Author Spotlight: A 3D Digital Model for the Diagnosis and Treatment of Pulmonary Nodules
Published on: May 19, 2023
2.4K
Conventional Filtering Versus U-Net Based Models for Pulmonary Nodule Segmentation in CT Images
Joana Rocha1,2, António Cunha3,4, Ana Maria Mendonça5,3
1Faculdade de Engenharia, Universidade do Porto, Rua Dr. Roberto Frias s/n, 4200 - 465, Porto, Portugal. joana866rocha@gmail.com.
Journal of Medical Systems
|March 7, 2020
Summary
Accurate lung nodule segmentation is crucial for early cancer diagnosis. Deep learning models like U-Net and SegU-Net show superior performance over conventional methods, aiding physicians in reliable lung pathology diagnosis.
Area of Science:
- Medical Imaging
- Artificial Intelligence
- Oncology
Background:
- Lung cancer diagnosis relies on accurate pulmonary nodule characterization.
- Segmentation of nodules in computed tomography (CT) scans is challenging due to variations in nodule appearance and surrounding anatomy.
- Improved nodule segmentation can enhance early detection and patient survival rates.
Purpose of the Study:
- To compare the effectiveness of different methodologies for pulmonary nodule segmentation in CT scans.
- To identify the most promising tool for improving nodule characterization and aiding lung cancer diagnosis.
- To evaluate a novel deep learning network, SegU-Net, against conventional and existing deep learning approaches.
Main Methods:
- Three segmentation methods were applied to 2653 nodules from the LIDC database: Sliding Band Filter (SBF), U-Net, and a novel SegU-Net.
- Performance was quantitatively assessed using the Dice score, comparing segmentation results against specialist-annotated ground truth.
- Computational cost and memory efficiency of the deep learning models were also evaluated.
Main Results:
- Deep learning models significantly outperformed the conventional SBF method, achieving higher Dice scores (U-Net: 0.830, SegU-Net: 0.823 vs. SBF: 0.663).
- U-Net and SegU-Net demonstrated high similarity to the ground truth, indicating reliable performance for nodule characterization.
- SegU-Net achieved comparable results to U-Net while offering reduced computational cost and improved memory efficiency.
Conclusions:
- Deep learning approaches, particularly U-Net and SegU-Net, are highly effective for pulmonary nodule segmentation in CT scans.
- SegU-Net presents a promising, efficient alternative for clinical implementation in decision support systems for lung cancer diagnosis.
- Accurate segmentation using these advanced methods can significantly assist physicians in establishing reliable diagnoses of lung pathologies.
