Can Data from One Medical Center be Enough to Generalize Lung Ultrasound Pattern Classification? A Multi-Center Domain Generalization Study

Domain shift refers to change of data distribution between training and testing datasets. In case of medical imaging, domain shift is extensive, specifically for multi-center studies. Different medical centers may use different scanners, imaging protocols, subject populations, etc. To mitigate this effect, domain generalization (DG) has been used over the time. In this regard, our focus is to analyze if a pre-trained model can generalize lung ultrasound (LUS) pattern classification among pneumonia patients. Furthermore, if LUS data from one medical center is enough to generalize classification task across different medical centers. Investigated LUS patterns include horizontal artifacts, vertical artifacts, and small to large consolidations. As a proof of concept, data from medical center in Brescia (Italy) is used as source dataset whereas data from medical center in Rome (Italy) is considered as target dataset. ResNet-18 model (pre-trained on ImageNet dataset) is employed. The pre-tra...

Domain shift refers to change of data distribution between training and testing datasets. In case of medical imaging, domain shift is extensive, specifically for multi-center studies. Different medical centers may use different scanners, imaging protocols, subject populations, etc. To mitigate this effect, domain generalization (DG) has been used over the time. In this regard, our focus is to analyze if a pre-trained model can generalize lung ultrasound (LUS) pattern classification among pneumonia patients. Furthermore, if LUS data from one medical center is enough to generalize classification task across different medical centers. Investigated LUS patterns include horizontal artifacts, vertical artifacts, and small to large consolidations. As a proof of concept, data from medical center in Brescia (Italy) is used as source dataset whereas data from medical center in Rome (Italy) is considered as target dataset. ResNet-18 model (pre-trained on ImageNet dataset) is employed. The pre-trained model is trained on the source dataset with transfer learning using linear probing (LP) and linear probing with fine-tuning (LP-FT) approach. Results show that the pre-trained model can generalize much better over the target dataset when it undergoes LP-FT rather than only LP, achieving a mean F1-Score of 63.08%. These findings encourage the use of pre-trained models to generalize across different medical centers for LUS data analysis. Furthermore, they suggest that one medical center as the source dataset may be enough to generalize across other medical centers with state-of-the-art comparable performance for LUS pattern classification.