Related Experiment Video
Updated: Jul 7, 2025
04:48
Application of Deep Learning-Based Medical Image Segmentation via Orbital Computed Tomography
Published on: November 30, 2022
2.8K
Enhancing COVID-19 CT Image Segmentation: A Comparative Study of Attention and Recurrence in UNet Models
Rossana Buongiorno1, Giulio Del Corso1, Danila Germanese1
1Institute of Information Science and Technologies, National Research Council of Italy (ISTI-CNR), 56124 Pisa, PI, Italy.
Journal of Imaging
|December 22, 2023
Summary
This study compared four AI models for segmenting COVID-19 lung infections on CT scans. Attention-UNet achieved the highest performance, demonstrating its effectiveness in disease stratification.
Area of Science:
- Artificial Intelligence in Medical Imaging
- Radiology and Medical Diagnostics
- Pulmonary Disease Analysis
Background:
- High-Resolution Computed Tomography (HRCT) is vital for COVID-19 management, aiding disease stratification and prognosis.
- Manual visual assessment of HRCT scans for COVID-19 abnormalities is time-consuming for radiologists.
- AI-driven semi-automatic segmentation tools can improve the efficiency and accuracy of detecting lung lesions.
Purpose of the Study:
- To compare the performance of four state-of-the-art Convolutional Neural Networks (CNNs) for the binary segmentation of COVID-19 infections in HRCT scans.
- To evaluate the effectiveness of incorporating attention mechanisms and recurrence paradigms into UNet architectures for this task.
- To understand the trade-offs between data requirements, computational resources, and model performance for COVID-19 lesion segmentation.
Main Methods:
- Trained and tested four encoder-decoder CNN models: UNet, Attention-UNet, R2-UNet, and R2-Attention UNet.
- Utilized 90 HRCT volumetric scans from COVID-19 patients, employing a five-fold cross-validation strategy.
- Assessed model performance using Dice score, Precision, and Recall, evaluated on both 2D images and 3D volumes.
Main Results:
- Attention-UNet demonstrated superior performance, achieving the highest 2D Dice score of 81.93% on the test set.
- Statistical analysis confirmed significant performance differences among the evaluated models.
- Integrating recurrence mechanisms (R2-UNet variants) did not improve, and potentially decreased, model effectiveness for this specific application.
Conclusions:
- Attention-UNet is a highly effective model for the semi-automatic segmentation of COVID-19 lung infections on HRCT scans.
- The study highlights the benefits of attention mechanisms in enhancing CNN performance for medical image analysis.
- Recurrence mechanisms may not be optimal for this particular COVID-19 segmentation task, suggesting careful architectural choices are necessary.
