Image-guided robotic Total Knee Arthroplasty (TKA) using Computed Tomography (CT) scans with the assistance of 3D femur models, has been proven to improve surgical outcomes as compared to traditional TKA procedures. However, The occurrence of bony projections or osteophytes at the knee joint significantly alters the anatomy, making surgical planning inaccurate. Problem Statement: Given an osteophytic 3D distal femur model, we propose an approach to remove osteophytes during pre-operative surgical planning. This facilitates accurate knee segmentation and precise kinematic computation, thus improving TKA outcomes. Method: We propose OsteoDeform, a geometric learning framework that applies continuous shape-based deformation flows, parameterized by a neural network, to the novel application of osteophyte removal in 3D femur models. Additionally, to address the lack of readily available datasets containing paired osteophytic and healthy distal femur models, we introduce the open-source OSTFemur-Dataset containing 495 unique models with paired correspondence, on which OsteoDeform is evaluated. Results: OsteoDeform effectively removes over \(70.7\%\) of distal femoral osteophytes. It also achieves a low Chamfer Distance loss of \(\sim \) 0.042 when compared with the corresponding healthy femur. A femur morphometric analysis was conducted, and OsteoDeform achieves an average morphometric error of \(\sim \) 0.022 \(\text {mm}\) , thereby preserving the anatomical integrity of the distal femur. We publicly release both the implementation of this research and the dataset: , .