A Machine Learning Based Algorithm to Determine Unloaded Geometry of the Breast Using MRI Image Data


  • Xi Feng Western University
  • Abbas Samani


Computer assisted medical procedures, Breast mechanical modeling, Reference geometry, Finite Element, Neural Networks


Biomechanical modelling has many medical applications in computer assisted diagnosis and intervention related to the breast. Examples include accurate breast cancer diagnosis, biopsies and surgeries, and breast post-surgery reconstruction. This approach also has industrial applications such as bra design. Breast mechanical models can be developed using Finite Element Method (FEM). Fundamental to reliable such breast models is the breast reference geometry under no loading. Most breast models use the breast Magnetic Resonance Image (MRI) to develop patient-specific FE models. However, the breast MRI scan is acquired under a prone body position which is associated with large breast tissue deformation resulting from gravity loading. As such, the breast MRI scans can only provide an approximate breast reference geometry, hence compromising the model’s expected accuracy. Such compromised accuracy can impact the accuracy of vital medical procedures such as breast biopsies that require needle targeting within few millimeters. In this study, an inverse algorithm is developed which aims at accurate determination of the breast reference geometry. In the proposed framework we generated two breast shape spaces, one filled with points representing a breast undeformed shape while the other containing points representing corresponding breasts deformed due to gravity loading under prone body position obtained using each breast’s FE model. To obtain a compact representation of the two spaces before fitting a function between them, principal component analysis was applied to each shape point set. A neural network was trained to find a mapping relationship between the two spaces. For validating the accuracy of reconstructed stress-free breast geometry, we applied gravityloading to assumed unloaded breast geometry using accurate FE simulation and used it as input geometry. To validate output stress-free breast shape, Intersection of Union (IoU) score and Hausdorff distance to compare it to the input breast geometry. Results indicated that the proposed inversion algorithm is accurate in capturing the breast’s stress-free configuration as well as in predicting its mechanical behavior.




How to Cite

X. Feng and A. Samani, “A Machine Learning Based Algorithm to Determine Unloaded Geometry of the Breast Using MRI Image Data”, CMBES Proc., vol. 46, Jun. 2024.



Clinical Engineering