Get to know PeekMed

Automatic landmark detection in 3D representation of orthopedic implants

July, 20 2023 1 minute read


Vítor Renato Pires Ferreira


Ana Costa & António Dourado Pereira Correia


Deep Learning, Pontos de referência, Artroplastia Total da Anca, Implantes, Redes Neuronais Convolucionais, Deep Learning, Landmarks, Total Hip Arthroplasty, Implants, Convolutional Neural Networks

Published on

Repositório Científico da Universidade de Coimbra


The planning of a surgery is essential for its success. One of the phases is the detection of anatomical landmarks in order to aid planning. With advances in software and medical imaging technologies, applications have emerged that can increase the accuracy of surgical planning. The process of manually detecting anatomical landmarks is time consuming and very dependent on the experience of the physician. With advances in Deep Learning (DL) techniques, namely Convolutional Neural Networks (CNN), this process can be automated. Total Hip Arthroplasty (THA) is known to provide reliable results, improving the patient’s quality of life. Implant placement is one of the steps in this procedure, and implant landmarks are needed to position the implant correctly. Usually, this process of manually annotating the implant landmarks is performed by specialists, involving detailed analysis of the implant geometry and marking the relevant points and axes. This is a time-consuming process, and a mechanism to automatically detect these landmarks is helpful. Just as CNNs can be used to detect anatomical landmarks, the process of detecting landmarks in orthopedic implants can be automated using CNNs. This thesis presents three approaches that use CNNs to automatically detect landmarks in stems, an orthopedic implant belonging to the hip implant system. Two of the three approaches, a CNN inspired by a scientific paper used to detect objects in dilated perivascular spaces, and a ResNet-34, use volumetric data as input data. The third approach, PointNet, uses point clouds as input data. Of all the approaches, PointNet was the best performing, with an Root Mean Square Error (RMSE) of 0.857mm ± 0.452mm, a relatively low error considering the average length of a stem.

Leave us your comments/doubts:
Date Article

Didn't find what you're looking for?


    Didn't find what you're looking for?