Time-domain and 3D methods for lung perfusion data clustering in electrical impedance tomography

The paper discusses the use of clustering methods to segment electrical impedance tomography images when assessing pulmonary perfusion through hypertonic saline injection. The clustering method assumes hybrid pixels (with both heart and lung partial volume effects) and lung pixels (solely due to lung perfusion). The study used data from 51 perfusions in 8 healthy and injured mechanically ventilated swine to generate ground truth masks and evaluate the clustering performance. Among the compared methods, k-means method with the correlation metric, was found to be the most effective and balanced performance. Achieving a median sensitivity and specificity of 84% while aiming to minimize false negative cases. This is important for avoiding the attribution of high lung perfusion values to voxels with mixed heart and lung effects.