Sequence Models of Artificial Intelligence for Pattern Recognition in Lung Ultrasound Videos

Lung ultrasound is a dynamic, non-invasive diagnostic tool that allows atelectasis to be observed after induction of general anesthesia. Recruitment maneuvers (RM) consist of transient increases in mean airway pressure to recruit collapsed alveolar units. The objective of the present study is to evaluate pulmonary aeration by utilizing ultrasound sequences captured over a period (video). We aim to recognize four different aeration patterns. Unlike previous approaches, in this case we consider the possible variation of patterns throughout the video, representing the sequence of frames as a sequence of feature vectors. To accomplish this, we employ the paradigm of Recurrent Neural Networks, specially designed for such cases, and integrate the study of Transformers, which also process temporal data. Due to the data limitation, values of Precision and Recall are considered preliminary, and their primary purpose is to validate the proof of concept presented in this work. They can be significantly enhanced by increasing the training data through a forward-looking procurement process. We anticipate that the successive studies we have been presenting will culminate in an automatic decision support system for monitoring pulmonary conditions.