RapidRead: Global Deployment of State-of-the-art Radiology AI for a Large Veterinary Teleradiology Practice

Abstract

This work describes the development and real-world deployment of a deep learning-based AI system for evaluating canine and feline radiographs across a broad range of findings and abnormalities. We describe a new semi-supervised learning approach that combines NLPderived labels with self-supervised training leveraging more than 2.5 million x-ray images. Finally we describe the clinical deployment of the model including system architecture, realtime performance evaluation and data drift detection.

Who Goes First? Influences of Human-AI Workflow on Decision Making in Clinical Imaging

“…When making diagnoses, participants were assisted by an AI diagnostic tool that estimated the likelihood of each finding being present in the X-ray image. See Fitzke et al [27] for a detailed description of the tool. The estimates are an average of predictions generated by eight separate convolutional neural networks, trained on a large proprietary dataset.…”

“…To convert the estimates produced by the AI tool into binary predictions, we adopted a threshold of 0.6 which corresponds to the choice made while deploying the model in production to help radiologists label new data for AI training purposes. This threshold was derived by calibrating each model to maximize its probability predictions at 0.5 with regards to Youden's J-Statistic [92] and then further calibrating the resulting ensemble to 0.6 due to its accuracy gains [27]. Participants had access to both the likelihood and the binary prediction of present versus absent generated by the AI tool for each finding, which in the remainder of the paper we call AI confidence and AI flags respectively.…”