Midline-Constrained Loss in the Anatomical Landmark Segmentation of 3D Liver Models

Abdul Karim Abbas, Aodhan Gallagher, Theodora Vraimakis, James Borgars, Ahmad Najmi Mohamad Shahir, Jibran Raja, Abhinav Ramakrishnan, Sharib Ali

Abstract: Anatomical landmark segmentation involves identifying specific points or regions within an anatomical structure and is integral to diagnostic processes and surgical guidance. This paper focuses on the segmentation of landmarks in 3D liver models in order to highlight key structures such as the falciform ligament and the liver ridge. The study is motivated by the need to support intraoperative laparoscopic registration tasks, aiming to enhance preoperative-to-intraoperative image fusion and thereby improving the localisation of tumours and vessels within the liver. As current practices typically rely on manual annotation, a process that is time-consuming and prone to human error when performed by less experienced operators, there is a clear need for a more efficient and reliable solution. Some recent works on landmark prediction in 3D liver models either over-predict or under-predict these landmarks. To overcome these challenges, we introduce a novel loss function that enforces geometric constraints by aligning segmentation predictions with a computed central anatomical midline. This strategy not only improves overall anatomical alignment but also ensures that the predictions remain thin and precise, reducing the occurrence of overly broad or misaligned outputs. This approach is utilised in conjunction with the PointNet++ architecture, trained on an extensive combined dataset composed of three smaller datasets, alongside the P2ILF challenge dataset, amounting to 300 unique 3D liver models in total. Our results indicate that our proposed solution forms a robust and precise approach, laying a solid foundation for future advances and feasibility in 3D-2D liver registration for intraoperative use. To improve reproducibility of this work, we have shared our code at: https://github.com/ARMADILLO-VISION/midline-loss.

External IDs:doi:10.1007/978-3-031-98691-8_16