Detecting the Absence of Lung Sliding in Lung Ultrasounds Using Deep Learning

Abstract

Certain post-thoracic surgery complications are monitored in a standard manner using methods that employ ionising radiation. A need to automatise the diagnostic procedure has now arisen following the clinical trial of a novel lung ultrasound examination procedure that can replace X-rays. Deep learning was used as a powerful tool for lung ultrasound analysis. We present a novel deep-learning method, automated M-mode classification, to detect the absence of lung sliding motion in lung ultrasound. Automated M-mode classification leverages semantic segmentation to select 2D slices across the temporal dimension of the video recording. These 2D slices are the input for a convolutional neural network, and the output of the neural network indicates the presence or absence of lung sliding in the given time slot. We aggregate the partial predictions over the entire video recording to determine whether the subject has developed post-surgery complications. With a 64-frame version of this architecture, we detected lung sliding on average with a balanced accuracy of 89%, sensitivity of 82%, and specificity of 92%. Automated M-mode classification is suitable for lung sliding detection from clinical lung ultrasound videos. Furthermore, in lung ultrasound videos, we recommend using time windows between 0.53 and 2.13 s for the classification of lung sliding motion followed by aggregation.

Explainability of deep learning models in medical video analysis: a survey

“…There are already practical deployments of AI, e.g. , to help doctors to identify the heart failure problems ( Choi et al, 2016 ), lung problems after thoracic surgery ( Jaščur et al, 2021 ) or automatic detection of COVID-19 from lung ultrasound ( Born et al, 2020 ). However, the full potential of AI systems is limited by the inability of the majority of algorithms to explain their results and decisions to human experts.…”

“…After clinical testing of a new procedure using lung ultrasound, the need arose to automate the diagnostic procedure. A study by Jaščur et al (2021) used DL in their work and created a new method that works with videos of lung ultrasound. The method consists of semantic segmentation of ultrasound images from the first images of the video.…”

Point-of-care AI-assisted stepwise ultrasound pneumothorax diagnosis

“…Recently, to overcome the variability of POC ultrasound, machine learning (ML) and deep learning (DL) have been intensively investigated, which involves various tasks including classification, segmentation, detection, registration, regression, and quality assessment for lung (Correa et al 2018, Short et al 2019, Sonko et al 2019, Baloescu et al 2020, Born et al 2020, Jascur et al 2021. In the domain of POC lung ultrasound, several DL methods have been developed to quantitatively assess pulmonary congestion by measuring the total B-line score (BLS).…”

“…Javsvcur et al developed CNN models trained on a relatively small dataset of 28 positive clips and 20 negative clips to classify the presence of lung sliding and perform pleura segmentation. They achieved an accuracy of 89% (Jascur et al 2021). Recent studies and developments have demonstrated the clinical feasibility of DL models in lung sliding classification, achieving AUCs over 90%, even with well-curated M-mode data (VanBerlo et al 2022, Zhang et al 2023).…”

“…The initial step in the clinician's lung ultrasound scan procedure involves locating an optimal position for the transducer on the chest, ensuring the pleural line is centered in the B-mode image. After determining the proper transducer position, the subsequent step involves capturing temporal M-mode images, as demonstrated in previous studies (Jascur et al 2021, VanBerlo et al 2022, Zhang et al 2023. By holding the transducer steady for 3-6 s, these images are utilized to assess the presence of lung sliding, with the barcode sign serving as a crucial indicator for PTX.…”