Learning-based landmark detection in pelvis X-Rays with attention mechanism: Data from the Osteoarthritis Initiative

We proposed a new landmark detection task based on the data from Osteoarthritis Initiative. We chose high-quality 544 images taken by the Swissray scanner from OAI bone image dataset. Images of the patients with hip replacement were excluded owing to the lack of anatomical structures that must be predicted, and some of the existing images that influenced the doctors labeling the landmarks were also excluded. In total, data of 524 subjects were collected to develop and validate a deep learning system to predict the landmarks in the pelvis images. According to our research problem, we invited professional doctors to label femoral head center, upper acetabulum rim, innermost point of subchondral sclerosis and teardrop in the pelvis X-Ray, as shown in the Figure 4. Total images were annotated by a radiologist with 7 years of working experience from the First Hospital of Jilin University, and the anatomical positions labeled were reviewed by orthopedic surgeon with 7 years of working experience from the Second Hospital of Jilin University.

There are two files in this dataset. "Coordinates_information.xlsx" stored the coordinates of landmarks labeled by radiologist. "original_information.xlsx" was download from Osteoarthritis Initiative, which stored the basic information about X-rays.