3D Shape Landmark Correspondence by Minimum Description Length and Local Linear Regularization

Abstract

Statistical Shape Models (SSMs) are currently used in orthopaedic surgery to allow accurate position of prosthetic components through bone morphing and to assess the correct post-operative follow up by virtually reconstructing the surgical site. Focusing on computer assisted Total Knee Arthroplasty (TKA) applications, in this paper we propose a new approach for establishing landmark correspondence of 3D shapes for building SSMs of anatomical structures around the knee joint. Our method is based on the landmark correspondence method byMinimum Description Length (MDL) and enforces local geometric similarity. Our new constraint, which is in the form of local linear regularization, ensures that the local shape geometry of corresponding landmarks on different shapes is similar. We tested our method on building SSMs of three anatomical structures from 24 MRI images of pathological knees, namely femur, patella and tibia. Compared with the original method using only the MDL criterion, our method shows significant improvement in two out of the three structures.