Deep learning and lung ultrasound for Covid-19 pneumonia detection and severity classification

Abstract

The Covid-19 European outbreak in February 2020 has challenged the world's health systems, eliciting an urgent need for effective and highly reliable diagnostic instruments to help medical personnel. Deep learning (DL) has been demonstrated to be useful for diagnosis using both computed tomography (CT) scans and chest X-rays (CXR), whereby the former typically yields more accurate results. However, the pivoting function of a CT scan during the pandemic presents several drawbacks, including high cost and cross-contamination problems. Radiation-free lung ultrasound (LUS) imaging, which requires high expertise and is thus being underutilised, has demonstrated a strong correlation with CT scan results and a high reliability in pneumonia detection even in the early stages. In this study, we developed a system based on modern DL methodologies in close collaboration with Fondazione IRCCS Policlinico San Matteo's Emergency Department (ED) of Pavia. Using a reliable dataset comprising ultrasound clips originating from linear and convex probes in 2908 frames from 450 hospitalised patients, we conducted an investigation into detecting Covid-19 patterns and ranking them considering two severity scales. This study differs from other research projects by its novel approach involving four and seven classes. Patients admitted to the ED underwent 12 LUS examinations in different chest parts, each evaluated according to standardised severity scales. We adopted residual convolutional neural networks (CNNs), transfer learning, and data augmentation techniques. Hence, employing methodological hyperparameter tuning, we produced state-of-the-art results meeting F1 score levels, averaged over the number of classes considered, exceeding 98%, and thereby manifesting stable measurements over precision and recall.

Review of Machine Learning in Lung Ultrasound in COVID-19 Pandemic

“…The knowledge of a pretrained DNN was transferred to new applications, and then only a few training parameters were needed, so the training dataset was largely reduced. Salvia et al employed residual CNNs (ResNet18 and ResNet50), transfer learning, and data augmentation techniques for multi-level pathology gradings (four main levels and seven sub-levels) over a dataset of 450 patients [ 105 ]. For both ResNet18 and ResNet50, and for both four-level and seven-level classifications, the metrics were all above 0.97.…”

“…Though many AI publications focused on binary classifications to differentiate COVID-19 from normal or other pneumonia cases, a few multiscale ML classifiers aimed to score stages of pneumonia and diagnose pathology severity quantitatively [ 62 , 78 , 79 , 80 , 81 , 82 , 88 , 93 , 98 , 105 ]. These severity scores could be used in the triage and management of patients in clinical settings.…”

Towards Accurate Point-of-Care Tests for Tuberculosis in Children

“…Obstacles comprise operator-dependency and lack of large data sets with standardized views. So far, AI endeavors for ultrasound imaging focused on algorithms for respiratory conditions because of the comparably consistent acquisition of images and homogenous sonographic pattern of healthy lungs [ 95 , 96 , 97 ]. Pilot data on a pediatric lung ultrasound-based AI algorithm, which correctly identified pneumonia infiltrates with 91% sensitivity and 100% specificity [ 97 ], indicate a potential future role for AI-based ultrasound image interpretation.…”

Automatic Deep Learning-Based Consolidation/Collapse Classification in Lung Ultrasound Images for COVID-19 Induced Pneumonia

“…The related literature on automatic diagnosis of COVID-19 using LUS images revolves around using Deep Learning (DL) algorithms trained on COVID-19 images [12][13][14][15] and on imaging patterns such as B-lines and pleural thickening that are associated with COVID-19 [16][17][18][19][20]. These DL algorithms consist of convolutional neural networks (CNNs), which are presently considered the state-of-theart for automated image analysis given their capability to extract low and high-level image features automatically.…”

“…These DL algorithms consist of convolutional neural networks (CNNs), which are presently considered the state-of-theart for automated image analysis given their capability to extract low and high-level image features automatically. La Salvia M et al [16] implemented a 4-(0-3) and a novel 7-class approach containing additional classes variations (namely 0, 0*, 1, 1*, 2, 2*, 3) to train a residual CNN where the classes vary from containing only Aline and B-lines (score 0) to artefacts resulting from the pleura and consolidated or tissue-like patterns (score 3). Alternatively, Arntfield et al [17] used B-lines from patients diagnosed either healthy, hydrostatic pulmonary oedema, or COVID to train a residual CNN similar to [16] to automatically detect COVID-19.…”