Supervised and unsupervised learning for lung perfusion data segmentation in electrical impedance tomography.

Abstract

Objective: Effective lung gas exchange relies on the balance between alveolar ventilation and perfusion, which can be disrupted in mechanically ventilated patients. Lung perfusion assessment using electrical impedance tomography (EIT) typically involves a sudden injection of a hypertonic saline solution. The large field of view provided by EIT often results in ambivalent behavior of many voxel waveforms following an indicator injection, where some exhibit indicator kinetics solely through the lungs (pulmonary), while others show passage through both the heart and lungs (hybrid). Consequently, a segmentation algorithm is essential for accurate perfusion evaluation.Approach: Sixteen pigs (29-35 kg) were mechanically ventilated and received a 10 ml bolus of 7.5% NaCl solution to assess lung perfusion during a healthy stage and, later, in an injured stage after receiving 3.5 ml kg^-1of HCl to induce acute lung injury. Supervised (Bagged Trees, Neural Networks, and Support Vector Machine) and unsupervised (K-means, Hierarchical, and Principal Component Analysis) learning methods were employed using 115 saline injections comprising voxel waveforms to label voxels as either hybrid or pulmonary. All segmentation methods were compared to a ground-truth mask manually drawn. A training dataset (81 injections) was used to train and cross-validate (five-fold) the supervised methods using previously extracted features. The test dataset (34 injections) was used to test both supervised and unsupervised learning algorithms.Main Results: A Principal Component Analysis (unsupervised learning) method exhibited the best overall performance, achieving 83% sensitivity, 92% specificity, 89% accuracy, and 84% dice similarity coefficient. No significant difference in performance was observed between healthy and injured subsets. Unsupervised methods consistently yielded more physiologically plausible and less scattered regions of interest.Significance: Accurate voxel labeling is crucial for lung perfusion assessment, as it enables discrimination of the indicator passage through the heart and lungs, thereby improving the estimation of regional pulmonary blood flow.