Tibial and femoral bones segmentation on CT-scans: a deep learning approach

Abstract

Custom implants in Total Knee Arthroplasty (TKA) could improve prosthesis’ durability and patient’s comfort, but designing such personalized implants requires a simplified and thus automatic workflow to be easily integrated in the clinical routine. A good knowledge of the shape of the patient's femur and tibia is necessary to design it, but segmentation is still today a key issue. We present here an automatic segmentation approach of the three joints of the lower limb: hip, knee and ankle, using convolutional neural networks (CNNs) on successive transverse views from CT images. Our three 2D CNNs are built on the U-net model, and their specialization each on one joint allowed us to achieve promising results presented here. This could be integrated in a TKA planning software allowing the automatic design of TKA custom implants.

nnU-Net for the Automatic Knee Segmentation from CT Images: A Comparative Study with a Conventional U-Net Model

“…Our 2D U-Net model has slightly changed since our previous publication [2], as the encoder now integrates a ResNet50V2 backbone pre-trained on ImageNet. However, the overall architecture and the fundamental outlines remain unchanged.…”

“…In a previous work [2], we have created a U-Net network [3] with an architecture which has proven to be fast and efficient for segmenting CT-Scans. We are now interested in comparing the performance of our model with a ready-made solution such as nnU-Net.…”

Intra-observer variations of femoral bony landmarks using three different methods for the design of custom knee implant

“…Fifty CT scans of lower limbs (right and/or left) from 30 patients were randomly selected from a CT database of the Brest University Hospital. They were automatically segmented using a U-Net model in order to obtain a 3D mesh of the distal femur 5 .…”

“…Those landmarks are used to calculate axes and (cutting) planes. A lot of literature is available on how accurate bone cuts are [2][3][4] but only a few studied the precision and variability of the landmarking process [5][6][7] .…”