Lightweight Transfer Learning Models for Ultrasound-Guided Classification of COVID-19 Patients

Abstract

Lightweight deep convolutional neural networks (CNNs) present a good solution to achieve fast and accurate image-guided diagnostic procedures of COVID-19 patients. Recently, advantages of portable Ultrasound (US) imaging such as simplicity and safe procedures have attracted many radiologists for scanning suspected COVID-19 cases. In this paper, a new framework of lightweight deep learning classifiers, namely COVID-LWNet is proposed to identify COVID-19 and pneumonia abnormalities in US images. Compared to traditional deep learning models, lightweight CNNs showed significant performance of real-time vision applications by using mobile devices with limited hardware resources. Four main lightweight deep learning models, namely MobileNets, ShuffleNets, MENet and MnasNet have been proposed to identify the health status of lungs using US images. Public image dataset (POCUS) was used to validate our proposed COVID-LWNet framework successfully. Three classes of infectious COVID-19, bacterial pneumonia, and the healthy lung were investigated in this study. The results showed that the performance of our proposed MnasNet classifier achieved the best accuracy score and shortest training time of 99.0% and 647.0 s, respectively. This paper demonstrates the feasibility of using our proposed COVID-LWNet framework as a new mobilebased radiological tool for clinical diagnosis of COVID-19 and other lung diseases.

Ultrasound-Based AI for COVID-19 Detection: A Comprehensive Review of Public and Private Lung Ultrasound Datasets and Studies

“…Therefore, AI studies often resize input images to a widely used common dimension across datasets. Most of the reviewed articles in this paper, for example, (Ebadi et al, 2021;Rojas-Azabache et al, 2021;Nabalamba, 2022;Quentin Muller et al, 2020;Karar et al, 2021a;Karnes et al, 2021;Perera et al, 2021), etc., also used the common image dimension of 224× 224 pixels, as well-known computer vision deep learning models are typically designed to intake images of 224×224 pixels. However, other image dimensions are also found for ultrasound COVID-19 studies.…”

Ultrasound-Based AI for COVID-19 Detection: A Comprehensive Review of Public and Private Lung Ultrasound Datasets and Studies

“…Therefore, AI studies often resize input images to a widely used common dimension across datasets. Most of the reviewed articles in this paper, for example, [37,49,50,[52][53][54][55], etc., also used the common image dimension of 224×224 pixels as well-known computer vision deep learning models are typically designed to intake images of 224×224 pixels. However, other image dimensions are also found for ultrasound COVID-19 studies.…”

“…In order to address the need for a less complex, power efficient, and less expensive solution to screen lung ultrasound images and monitor lung status, Hou et al [75] introduced a Saab transform-based subspace learning model to find the A-line, B-line, and consolidation in lung ultrasound data. Karar et al [53] introduced a lightweight deep model, COVID-LWNet, to make an efficient CNN-based system for classifying lung ultrasound images into COVID-19, bacterial pneumonia, and healthy classes. In addition, Karar et al [56] proposed a generative adversarial network (GAN) to perform the same task on ultrasound images.…”

Ultrasound-Based AI for COVID-19 Detection: A Comprehensive Review of Public and Private Lung Ultrasound Datasets and Studies

“…Therefore, AI studies often resize input images to a widely used common dimension across datasets. Most of the reviewed articles in this paper, for example, [37,49,50,[52][53][54][55], etc., also used the common image dimension of 224×224 pixels as well-known computer vision deep learning models are typically designed to intake images of 224×224 pixels. However, other image dimensions are also found for ultrasound COVID-19 studies.…”

“…We also describe different types of AI models, used by state-of-the-art US COVID-19 studies in the following sections. [59] SqueezeNet, MobileNetV2 ✗ Al-Jumaili et al [68] ResNet-18, RestNet-50, NASNetMobile, GoogleNet, SVM Al-Zogbi et al [70] DenseNet ✗ Almeida et al [71] MobileNet ✗ Arntfield et al [38] Xception ✗ Awasthi et al [72] MiniCOVIDNet ✗ Azimi et al [73] InceptionV3, RNN ✗ Barros et al [69] Xception-LSTM ✗ Born et al [12] VGG-16 ✗ Born et al [74] VGG-16 ✗ Born et al [13] VGG-16 ✗ Carrer et al [16] Hidden Markov Model, Viterbi Algorithm, SVM ✗ Che et al [17] Multi-scale Residual CNN ✗ Chen et al [40] 2-layer NN, SVM, Decision tree Diaz-Escobar et al [67] InceptionV3, VGG-19, ResNet-50, Xception ✗ Dastider et al [18] Autoencoder-based Hybrid CNN-LSTM ✗ Durrani et al [35] Reg-STN ✗ Ebadi et al [52] Kinetics-I3D ✗ Frank et al [19] ResNet-18, MobileNetV2, DeepLabV3++ ✗ Gare et al [15] Reverse Transfer Learning on UNet ✗ Hou et al [75] Saab transform-based SSL, CNN ✗ Huang et al [41] Non-local channel attention ResNet ✗ Karar et al [53] MobileNet, ShuffleNet, MENet, MnasNet ✗ Karar et al [56] A semi-supervised GAN, a modified AC-GAN ✗ Karnes et al [54] Few-shot learning using MobileNet ✗ Khan et al [76] CNN ✗ La Salvia et al [42] ResNet-18, ResNet-50 ✗ Liu et al [48] Multi-symptom multi-label (MSML) network ✗ MacLean et al [77] COVID-Net US ✗ MacLean et al [78] ResNet ✗ Mento et al [44] STN, U-Net, DeepLabV3+ ✗ Muhammad and Hossain [58] CNN ✗ Nabalamba [49] VGG-16, VGG-19, ResNet ✗ Panicker et al [36] LUSNet (a U-Net like network for ultrasound...…”