Landmark prediction in large radiographs using RoI-based label augmentation

In this work, we present a novel approach to automate landmark detection in full length leg radiographs (“long leg films”), as used in pre- and post-surgical assessment of total knee arthroplasty. We show experimentally that single step methods only using a U-Net model struggle to achieve high-accuracy landmark predictions for large radiograph images. We instead propose a two-step system that combines a recurrent-convolutional neural network (R-CNN) with a U-Net model enhanced by a dilation-erosion scheme that initially dilates labels to create larger regions and then progressively erodes them back to their original sizes over training epochs for precise landmark detection. We demonstrate the significant improvements in accuracy and reliability that this scheme provides over baseline method using a dataset of long leg films with high-standard annotations.