Proximal Femur Image Database Validation

The AO@AI Turin project is a collaborative project with a Turin group and the AO (Arbeitsgemeinschaft für Osteosynthesefragen, or in English, Association for the Study of Internal Fixation) foundation. An Image database (DB) has been built to host AP pelvic radiographs ready for artificial intelligence (AI) development. The goal of this project is to determine the agreement between the Turin annotation of fracture status and the annotation from an external group of AO expert surgeons for a random subset of the Turin images.

The AO@AI Turin project is a collaborative project with a Turin group who has collected 2,932 anteroposterior (AP) pelvic radiographs, of which 1,811 are fracture images, and 1,121 are non-fracture images. The Turin group has developed an artificial intelligence (AI) algorithm for fracture classification using these images. These anonymized images (with all metadata or personal identifiers removed) have been uploaded to a cloud-based image database (DB) hosted and managed by the AO Foundation. The Turin group has established the "ground truth" using the methods of "consensus by experts". Two radiologists from their medical team have reviewed and classified the fracture status (fracture vs non-fracture, and, if fracture, the AO/Orthopedic Trauma Association \[OTA\] classification). The next step's goal is the ground truth validation plan to test the accuracy of the Turin annotation of fracture classification of the already uploaded AP pelvic images. This is to ensure that the image DB offers accurate quality annotations to allow AI development. For the pilot phase, a random subset of the Turin images (300 of images) will be drawn from the image DB. These images will be reviewed by an external group of AO expert surgeons who will annotate the images per their fracture status, i.e., fracture vs non-fracture, and, if fracture, the AO/OTA classification. The group of AO expert surgeons consists of four surgeons who will independently review the 300 images and a fifth surgeon who serves as an adjudicator if necessary. The expert surgeons will be given access to the 300 images via the cloud-based image DB and annotate the images. The expert surgeons will be blinded to the Turin annotations. The expert surgeons' annotations will be entered into a DB built for the purpose for the pilot study. To determine the ground truth, the annotations of the four surgeons will be compared, and discrepancies will be identified. A meeting will then be arranged among the surgeons to resolve, by consensus, the discrepancies, with the potential involvement of the fifth surgeon as the adjudicator. After the resolution meeting, there will be a single set of annotations for the 300 images from the exert surgeon group. The Turin annotations will also be entered into the study DB to allow comparisons with the expert surgeon group's annotation. In case of disagreement between the Turin annotation and the AO expert surgeon annotations, a consensus will be sought to establish a new ground truth. If this process results in significant revisions to the annotations, the entire dataset will be reviewed to set this new standard. Following such a comprehensive dataset revision, the algorithm for automated fracture classification of the proximal femur, which has already been developed by the Turin group, will be re-trained. After re-training, the algorithm's performance will be evaluated through metrics such as precision, recall, and F1-score to ensure its accuracy and effectiveness in classifying proximal femur fractures.