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.

Keywords: Deep learning; LUS Score; Lung ultrasound; SARS-CoV-2.

Graphical abstract

Fig. 1

Percentage distribution of frames for each classification task. Left: four class scenario; right: seven class scenario. The percentage of images assigned to each score for both diagnostic tasks is depicted; pleural line involvement is highly likely and more severe when a frame is assigned a high score.

Fig. 2

Examples of selected and rejected frames.

Fig. 3

Augmented training set images: augmentations described in this section have been applied to the training images and are shown in this figure.

Fig. 4

– Residual Network Structure Diagrams: plot of each ResNet employed together with their structure and exploited layers.

Fig. 5

- Inference scalability: processing times [s] according to batch size.

Fig. 6

ResNet50 Class Activation Mapping, seven class scenario: both severity scoring, B-lines and pleural line consolidations and irregularities are correctly highlighted along with tissue-like patterns for Score 3.