Lung ultrasound: ZAGNet analyses each zone individually, as if it were a graph, and improves the diagnosis of consolidation by up to 19 per cent

Rather than summarising dozens of clips with a simple maximum or average value, this network links neighbouring lung areas and accommodates incomplete examinations.

A lung ultrasound examination scans several areas of the chest, but in practice some are often skipped. Existing AIs mainly analyze isolated images or clips, then aggregate their results in a rudimentary way, by maximum or average, without taking into account the anatomical proximity between the regions. A team of researchers proposes ZAGNet, a graph-based neural network where each anomaly tracked over time becomes a node, connected to the others according to the adjacency of the lung areas.

Net gains on 714 patients

A graph transformer circulates information between neighboring regions, and a global node produces a patient-wide diagnosis for lung condensation and pleural effusion, with only one tag per patient for training. Because the graph does not have a fixed size, missing areas are not a problem. On a multicenter set of 714 subjects and 20,256 clips, with examinations of 4 to 16 areas, the area under the curve increases to 0.803 for condensation (compared to about 0.68 for conventional aggregations) and 0.893 for effusion (compared to 0.80 to 0.82), i.e. gains of up to 19% and 11%.