PREDICTING MISSING BONE ANATOMY: HOW GOOD CAN WE GET?

A variety of imaging modalities and strategies are deployed to capture sufficient detail for extracting key anatomical parameters used in diagnostics and research. Imaging protocols are often clinic-specific, with budget and time limitations, and clinical priorities. This can result in a loss of additional anatomical measures that could prove useful in future retrospective applications of the images. Statistical shape modelling is a robust tool for predicting missing bone shape from partial medical imaging and automating anatomical measure extraction. We demonstrate a pipeline to predict missing bone shape from limited field-of-view MRI and CT imaging to automatically calculate anatomical measures that were otherwise not visible in the original scan. We present the 2-stage validation of a femur SSM built from 50 healthy cadaveric CT. The model was used to predict femur shapes using two different fitting tasks; using data from the full femur, and using data limited to the proximal 15cm of the femur, simulating a limited field-of-view scan. The correspondence generated from the fitting task was used to automatically extract five key anatomical measures. In the first validation step the model was used to reconstruct the Training dataset (50 healthy femurs used to build the model) and a Test dataset (17 healthy femurs excluded from the model-building process) from both full and limited field-of-view settings to test the robustness of the model with healthy data input to predict missing bone parts. In the second validation step the model was fit to a limited field-of-view clinical MRI dataset (23 femurs with DDH) and the predicted anatomical measures were compared to manual measures taken from clinical 2D x-ray imaging, using the measures from the MRI as a ground truth. There were no main or interaction effects for error in validation step 1, but differences were found for Hausdorff Distance and Dice Score. In validation step 2, no significant differences were found between the error of the anatomical measures of the model prediction and x-ray method. The pipeline presents a promising method for automating diagnostic measurements of femur anatomy, from both full and limited field-of-view imaging, in good agreement with the literature.