PhD opportunities

[Subject PROVIDED] [Transfer learning in biomechanics] High-dimensional transfer learning for personalized biomechanical modeling in surgery planning: application to anterior-cruciate ligament reconstruction

Thesis proposal

Area of expertiseMécanique
Doctoral SchoolISMME - Ingénierie des Systèmes, Matériaux, Mécanique, Énergétique
SupervisorRYCKELYNCK David
Co-supervisorCORTE Laurent
Research unitCentre des Matériaux
Keywords/, /
AbstractNumerical modeling offers unprecedented possibilities for surgery planning by allowing quantitative patient-specific predictions, therefore reducing the risks of complications and improving the safety of the patients [3]. While the first generation of models are solely based on bioimaging data, numerous surgical problems necessitate a more complex description that takes into account the mechanics and kinematics of the tissues and implants of interest [4, 5], with patient-specific images as input data. The reconstruction of the Anterior Cruciate Ligament (ACL) in Fig. 1a., which is the most frequent ligament injury (incidence rate 1/3000), is an excellent illustration of this challenge. Recent works have shown that friction against the cartilage of the femoral notch and condyle is a major cause of failure for reconstructed ACL. The occurrence and intensity of friction is highly sensitive to the location of the femoral and tibial insertion points chosen by the surgeon (Fig. 1b) and to the local bones shape. Such contact problem is extremely dificult to anticipate as it depends not only on the geometry but also on the kinematics of each knee (Fig. 1c). Today, the image-based biomechanical modeling procedure is far away too slow for any practical application to patient-specific modeling in view of surgery planning.
Our project proposes to tackle this issue by associating Artifficial Intelligence approaches, model reduction and image-based biomechanical modeling to develop efficient numerical models for ACL reconstruction. The objective is three-fold: (i) to provide a clinically-relevant model applicable to fast surgical planning; (ii) to provide a deeper understanding of the damaging mechanisms of reconstructed ligaments; (iii) to develop deep transfer learning methods for image-based modeling in biomechanics [6, 7, 8, 9]. For that, input data coming from different sources will be used: radiographic data [10, 11], magnetic resonance images [12, 13], experimental data on animal models [2], data on the variability of surgeon accuracy.

The project will benefit from an established collaboration with the consortium of the LIGAGEL project [2], which aims at developing novel artificial ligaments for ACL reconstruction [14]. This will provide access to complete data sets including 3D biomechanical simulations, 3D imaging and kinematics for both sheep (see for instance Fig.1b-c) and Human knee joints as well as post-operative observations of ligament and cartilage damage in the case of animal studies. This project will be a success if patient specific digital images and hyperelastic data, with different source domains, are merged in a fast image-based modeling chain, with a simulation speed up of 100. In high-dimensional transfer learning, dimensionality reduction is one of the most important ways to preserve the discriminant information for subsequent classification [15] or for model order reduction via a ROM-net[16]. In 2 reduced order models, both deep learning and physical equations are coupled in a single modeling procedure. It achieves transfer learning. The main advantage of transfer learning is its ability to reuse data related to various source domains, here biomechanics and image classification, when it is expensive or impossible to re-collect the needed training data for the target domain [6, 7, 8, 9]. Recent advances in U-nets [17, 18] and in multimodal autoencoders [19] that extract a common latent space from various sources of data, will certainly foster fast image-based reduced predictions with transfer learning. A particular attention will be given to the description of the sensitivity to the femoral and tibial insertion points, to the bones geometry and to the mechanical behavior of the ligament substitute.
The PhD student will join the \Artificial Intelligence for the Sciences' training program. All observational data and numerical being already available, we are confident that several papers will be published in 3 years. More over, we already have a good experience in U-nets [18] and in transfer learning via ROM-nets [16] for mechanical modeling.
FundingFinancement d'un Etablissement d'enseignement supérieur