Related Experiment Video
Updated: Aug 8, 2025
04:48
Application of Deep Learning-Based Medical Image Segmentation via Orbital Computed Tomography
Published on: November 30, 2022
2.8K
DEHA-Net: A Dual-Encoder-Based Hard Attention Network with an Adaptive ROI Mechanism for Lung Nodule Segmentation.
Muhammad Usman1, Yeong-Gil Shin1
1Department of Computer Science and Engineering, Seoul National University, 1 Gwanak-ro, Gwanak-gu, Seoul 08826, Republic of Korea.
Sensors (Basel, Switzerland)
|February 28, 2023
Summary
Accurate pulmonary nodule segmentation aids early lung cancer diagnosis. This novel 3D method uses a dual-encoder hard attention network (DEHA-Net) and an adaptive region of interest (A-ROI) algorithm for improved segmentation accuracy.
Area of Science:
- Medical Imaging
- Artificial Intelligence in Healthcare
- Radiology and Oncology
Background:
- Accurate pulmonary nodule segmentation is crucial for early lung cancer diagnosis and improving patient survival rates.
- Existing segmentation methods often rely on manual 3D volumetric or fixed 2D regions of interest, limiting detection and potentially causing inaccuracies.
- These limitations hinder the ability to detect nodules outside predefined areas and can include irrelevant structures in segmentation.
Purpose of the Study:
- To propose a novel approach for 3D lung nodule segmentation that overcomes the limitations of existing methods.
- To develop a two-stage segmentation technique utilizing a dual-encoder-based hard attention network (DEHA-Net) and an adaptive region of interest (A-ROI) algorithm.
- To improve the accuracy and robustness of lung nodule segmentation for enhanced early lung cancer detection.
Main Methods:
- A two-stage segmentation process was developed, starting with DEHA-Net using axial CT slices and an initial 2D region of interest (ROI).
- An adaptive region of interest (A-ROI) algorithm automatically generated ROIs for surrounding slices, eliminating the need for further radiologist input.
- The second stage employed DEHA-Net on sagittal and coronal views, with a consensus module integrating all segmentations for the final 3D volumetric result.
Main Results:
- The proposed method achieved an average Dice score of 87.91%, sensitivity of 90.84%, and positive predictive value of 89.56% on the LIDC/IDRI dataset.
- The framework demonstrated significant robustness across various nodule types, shapes, and dimensions.
- Quantitative analysis confirmed improved performance compared to existing state-of-the-art lung nodule segmentation methods.
Conclusions:
- The novel two-stage 3D lung nodule segmentation approach significantly enhances segmentation accuracy and robustness.
- The integration of DEHA-Net and A-ROI effectively addresses limitations of previous methods, improving early lung cancer diagnosis potential.
- The method shows promise for more reliable and automated lung nodule detection in clinical settings.
