Exploring Global Clinical Translation: Cross-Validation of Institutional-Specific AI Models for Lung Ultrasound Diagnosis

Abstract

After the pandemic, various deep learning models have been developed to classify lung ultrasound patterns. Most institutions train their models on local datasets. Some gather data from hospitals where clinicians use a variety of probes, while others standardize the process by requiring a specific probe. Cross-institutional validation of these models on datasets from international institutions is infrequent. This study aims to evaluate whether models developed with different approaches in separate institutions could achieve comparable performance on external datasets. Two models previously developed by two different international institutions were cross-validated. The first model is a classification model (CM) developed by ULTRa lab from Italy. The second model is a segmentation model (SM) with signal post-processing algorithms trained at institution CSIC from Spain. We evaluated the performance of models on two external test datasets at exam and prognostic level. CM and SM models achieved comparable results. These findings suggest that models trained individually in different institutions with different approaches can still achieve comparable results, highlighting the potential for multi-center clinical translations.

Ähnliche Arbeiten