LIZARD - Liver resection zone prediction using image-based and geometric deep learning

• Applicants: Dr. Florentine Huettl; Professorin Dr.-Ing. Sylvia Saalfeld
• Subject Area: Medical Informatics and Medical Bioinformatics
• Term: since 2024
• Project identifier: Deutsche Forschungsgemeinschaft (DFG) - Project number 547369510

Following cardiovascular diseases, cancer constitutes the second major cause of death, with the liver being one of the most common sites for the development of primary and metastatic lesions. Due to the complexity of the liver anatomy and its high interindividual anatomical variability, liver surgery requires patient-specific careful preoperative planning depending on the location and extent of the tumor mass. For surgical planning, 3D reconstructions of the liver, the feeding and draining vessels, the tumor, and the bile ducts, are created from the tomographic CT or MRI data sets. In addition, the resection plan around the tumor is often created manually or semi-automatic, which is time-consuming and user-dependent. The LIZARD project aims to utilize deep learning-based segmentation, mesh generation, and mesh deformation to achieve patient-specific resection planning for liver surgery. A very import part of the project is the compilation of a training database, which will furthermore enable liver anatomy classification. In addition, clinical data will be integrated to enhance the image-based methods. Thus, the project will be able to consider the tumor entity and will be systematically expanded to patients with multiple lesions. This proposal focuses on developing adapted deep learning strategies that enable liver surgeons and clinical researchers to predict liver resection volume, therefore, the accuracy and interpretability of the trained networks will be controlled and enhanced. Since this interdisciplinary research project involves computer scientists and medical doctors, i.e., radiologists and surgeons, the interpretability of the trained networks is crucial. In addition, the determination of areas of interest of the trained networks and analysis of anatomical regions within the geometric deep learning approaches will be carried out. Finally, a critical clinical evaluation will be conducted and the expert feedback will be integrated. By using the proposed deep learning strategies, we aim for an easy-to-access and fast solution that does not require (once it is trained) expensive hardware or highly trained professionals. As a result, the LIZARD project allows for the assessment of the grading of the patient and assignment to specialized centers based on the complexity of resection due to tumor state, localization and size. Thus, more patients can undergo a potentially curative resection.

• DFG Programme: Research Grants