Arthroscope Localization in 3D Ultrasound Volumes Using Weakly Supervised Deep Learning

This work presents an algorithm based on weak supervision to automatically localize arthroscope 3D ultrasound (US). The ultimate goal of this application is combine US with the 2D view during knee arthroscopy, provide surgeon a comprehensive surgical site. implemented consisted weakly supervised neural network, which was trained images different phantoms mimicking imaging conditions arthroscopy. Image-based classification performed and resulting class activation maps were used arthroscope. localization performance evaluated visually by three expert reviewers calculation objective metrics. Finally, also tested human cadaver knee. achieved average accuracy 88.6% phantom data 83.3% data. correct in 92–100% all true positive classifications for both These results are relevant because they show feasibility automatic volumes, paramount combining multiple image modalities that available arthroscopies.