Learning the Imaging Landmarks: Unsupervised Key point Detection in Lung Ultrasound Videos
This work addresses the need for data-driven landmark identification in lung ultrasound to improve diagnostic accuracy for COVID-19 patients, representing an incremental advance in medical imaging.
The paper tackled the problem of automatically detecting key landmarks in lung ultrasound videos for COVID-19 monitoring by developing an unsupervised method using transporter neural networks, achieving 91.8% accuracy on pleura detection with 1081 video frames.
Lung ultrasound (LUS) is an increasingly popular diagnostic imaging modality for continuous and periodic monitoring of lung infection, given its advantages of non-invasiveness, non-ionizing nature, portability and easy disinfection. The major landmarks assessed by clinicians for triaging using LUS are pleura, A and B lines. There have been many efforts for the automatic detection of these landmarks. However, restricting to a few pre-defined landmarks may not reveal the actual imaging biomarkers particularly in case of new pathologies like COVID-19. Rather, the identification of key landmarks should be driven by data given the availability of a plethora of neural network algorithms. This work is a first of its kind attempt towards unsupervised detection of the key LUS landmarks in LUS videos of COVID-19 subjects during various stages of infection. We adapted the relatively newer approach of transporter neural networks to automatically mark and track pleura, A and B lines based on their periodic motion and relatively stable appearance in the videos. Initial results on unsupervised pleura detection show an accuracy of 91.8% employing 1081 LUS video frames.