MedmeshCNN - Enabling meshcnn for medical surface models. (October 2021)

• Record Type: Journal Article
• Authors: Schneider, Lisa; Niemann, Annika; Beuing, Oliver; Preim, Bernhard; Saalfeld, Sylvia
• Is Part Of: Computer methods and programs in biomedicine. Volume 210(2021)
• Publication Date: 2021-10
• Subjects: Geometric deep learning -- Mesh processing -- Shape segmentation -- Intracranial aneurysms -- Surface models -- Convolutional neural network; Medicine -- Computer programs -- Periodicals; Biology -- Computer programs -- Periodicals; Computers -- Periodicals; Medicine -- Periodicals; Médecine -- Logiciels -- Périodiques; Biologie -- Logiciels -- Périodiques; Biology -- Computer programs; Medicine -- Computer programs; Periodicals; Electronic journals; 610.28
• DOI: 10.1016/j.cmpb.2021.106372
• Languages: English
• ISSNs: 0169-2607

Abstract

Highlights: MedMeshCNN enables the application of MeshCNN on 3D medical surface models. Improved computational efficiency allows retaining fine-grained patterns and patient-specific properties. Weighted loss function enables the processing of highly imbalanced datasets with pathological findings. Manifold assumption of MeshCNN is relaxed and meaningful holes within meshes are leveraged as features. MedMeshCNN keeps pace with the MeshCNN performance under significantly more complex conditions. Abstract: Background and objective: MeshCNN is a recently proposed Deep Learning framework that drew attention due to its direct operation on irregular, non-uniform 3D meshes. It outperformed state-of-the-art methods in classification and segmentation tasks of popular benchmarking datasets. The medical domain provides a large amount of complex 3D surface models that may benefit from processing with MeshCNN. However, several limitations prevent outstanding performances on highly diverse medical surface models. Within this work, we propose MedMeshCNN as an expansion dedicated to complex, diverse, and fine-grained medical data. Methods: MedMeshCNN follows the functionality of MeshCNN with a significantly increased memory efficiency that allows retaining patient-specific properties during processing. Furthermore, it enables the segmentation of pathological structures that often come with highly imbalanced class distributions. Results: MedMeshCNN achieved an Intersection over Union of 63.24% on a highly complex part segmentation task of intracranial aneurysms and their surrounding vessel structures. Pathological aneurysms were segmented with an Intersection over Union of 71.4%. Conclusions: MedMeshCNN enables the application of MeshCNN on complex, fine-grained medical surface meshes. It considers imbalanced class distributions derived from pathological findings and retains patient-specific properties during processing.