Validation of deep learning enabled software MetronMind to measure vertebral heart size and vertebral left atrial size in dogs

Abstract

Background Vertebral heart size (VHS) and vertebral left atrial size (VLAS) are objective radiographic measurements of heart and left atrial size, respectively and are associated with inter- and intraobserver variability when measured by humans. Artificial intelligence (AI) tools to determine VHS and VLAS have been developed and may reduce variability and save time. Objectives To compare two manual methods for measuring VHS and VLAS on right and left lateral canine thoracic radiographs and to compare measurements of VHS and VLAS made by deep learning-enabled program, MetronMind, with those made by a trained observer on right and left lateral radiographs. Animals Client-owned dogs (n = 1058) including 80 breeds with a variety of heart sizes, thoracic conformations and radiographic quality. Methods Retrospective, single-center, method-comparison study. Pearson’s correlation, Bland-Altman plots and Passing-Bablok regression were used to assess agreement. Results Correlation between traditional and modified manual measurements for VHS and VLAS was strong (r = 0.994 and r = 0.974, respectively), with minimal bias (−0.10 and 0.04 vertebrae, respectively) indicating that the modified methods closely approximate traditional measurements obtained from right lateral views. MetronMind measurements of VHS and VLAS from right lateral radiographs also correlated well with the human observer’s modified measurements (r = 0.947 and r = 0.811 respectively), showing small mean biases (0.08 and 0.07 vertebrae, respectively). Correlation between left and right lateral radiographic measurements of VHS (0.87 and 0.91) was higher than for VLAS (0.73 and 0.64) and bias was larger for VHS (0.26 and 0.31 vertebrae) than for VLAS (−0.13 and −0.10 vertebrae) for humans and MetronMind, respectively.