LwMLA-NET: A Lightweight Multi-Level Attention-Based NETwork for Segmentation of COVID-19 Lungs Abnormalities From CT Images
- 23 March 2022
- journal article
- research article
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Instrumentation and Measurement
- Vol. 71 (00189456) , 1-13
Abstract
COronaVIrus Disease 2019 (COVID-19) emerged as a global pandemic in the last two years. Typical abnormal findings in chest computed tomography (CT) images of COVID-19 patients are ground-glass opacities (GGOs) and consolidation, which signify the extent of damage caused to the lungs. The manual annotation of these abnormalities for severity analysis is complex, tedious, and time-consuming. This motivated us to develop a vision-based analysis framework for automated segmentation of lung abnormalities. We proposed a deep learning framework, namely LwMLA-NET “Lightweight Multi-Level Attention-based NETwork,” to segment GGO and consolidation. The LwMLA-NET is based on an encoder–decoder architecture where depth-wise separable convolutions are employed at each stage, making it a light-weighted framework that significantly reduces the computational cost. Another distinguishable module in LwMLA-NET is the multilevel attention (MLA) mechanism that focuses on dominant and relevant features and avoids propagation of insignificant features from the encoder to the decoder, thereby aiding faster optimization. Integrating the atrous spatial pyramid pooling (ASPP) module in the bottleneck helps to handle scale variations. The LwMLA-NET was evaluated on two databases—MedSeg and Radiopedia—and obtained F1-scores of 76.7% and 73.1%, respectively. The experimental evaluation justifies that LwMLA-NET outperforms other state-of-the-art deep learning frameworks like Attention U-Net, PSP-Net, Cople-Net, Inf-Net, and Mobile-Net V2 in terms of segmentation performance and has an acceptable generalization capability. Moreover, our team, “LwMLA-NET-Team-KDD-JU,” participated in Kaggle’s “COVID-19 CT Images Segmentation” open challenge. The performance was evaluated on a separate test set, and we obtained fourth rank on the leaderboard among 40 teams with an F1-score of 71.196%.Keywords
Funding Information
- Project IT4Neuro degeneration (CZ.02.1.01/0.0/0.0/18 069/0010054)
- University of Hradec Kralove Faculty of Informatics and Management Czech Republic (2021/2204)
- Grant Agency of Excellence
- University of Hradec Kralove
- Faculty of Informatics and Management Czech Republic
This publication has 39 references indexed in Scilit:
- Generalised Dice Overlap as a Deep Learning Loss Function for Highly Unbalanced Segmentations2017
- Xception: Deep Learning with Depthwise Separable Convolutions2017
- Pyramid Scene Parsing Network2017
- DeepLab: Semantic Image Segmentation with Deep Convolutional Nets, Atrous Convolution, and Fully Connected CRFsIEEE Transactions on Pattern Analysis and Machine Intelligence, 2017
- Localized Energy-Based Normalization of Medical Images: Application to Chest RadiographyIEEE Transactions on Medical Imaging, 2015
- Automated Vessel Segmentation Using Infinite Perimeter Active Contour Model with Hybrid Region Information with Application to Retinal ImagesIEEE Transactions on Medical Imaging, 2015
- Camera as the instrument: the rising trend of vision based measurementIEEE Instrumentation & Measurement Magazine, 2014
- Metrological Characterization of a CADx System for the Classification of Breast Masses in MammogramsIEEE Transactions on Instrumentation and Measurement, 2010
- Quantitative RT-PCR: Pitfalls and PotentialBioTechniques, 1999
- A feature-integration theory of attentionCognitive Psychology, 1980