Evaluation of The Use of a Deep Active Learning Model in Anatomic Segmentation Tasks in Canine Thoracic Radiographs

Journal Title
Journal ISSN
Volume Title
University of Guelph

The main objective of this work was to assess the use of a semi-automated segmentation method compared with a manual method for canine thoracic radiographs. Additionally, this study seeks to compare the accuracy of novice evaluators to that of experts in basic anatomic segmentation tasks using an active learning (AL) model. The artificial intelligence algorithm was trained using 900 thoracic radiographs from patients referred to the Ontario Veterinary College between January 2020 to July 2021. Participants achieved better intersection over union (IoU) and Hausdorff distance scores using the semi-automatic method for the segmentation of the heart, abdomen, and spinous process in comparison to the manual method. There were no significant differences in the mean IoU scores between cohorts for the automatic method. In conclusion, the AL segmentation model is a feasible model for assisting users with varying levels of radiology experience to segment anatomical structures effectively on canine thoracic radiographs.

Machine Learning, Thoracic Radiographs, Deep Learning, Veterinary Diagnostic Imaging