Deep Learning Methods in 3D Computed Tomography Images for Implantable Devices

Implantable medical devices have revolutionized modern medicine, offering life-saving interventions and significantly improving the quality of life for many patients. The process of implanting a medical device involves several key stages to ensure patient safety and achieve optimal results. These personalized stages include medical evaluation, diagnosis, therapy selection, surgical planning, surgical procedure, and postoperative follow-up.
This thesis focuses on developing advanced 3D image analysis methods to enhance preoperative tasks in CT images involving procedures for medical implantable devices. More specifically for cochlear implantation and left atrial appendage occlusion, these procedures involve navigating complex and variable anatomical structures, where accurate diagnosis and planning are crucial for successful outcomes.
A significant part of the work is dedicated to cochlear implantation, where the intricate anatomyof the inner ear presents substantial challenges. The thesis introduces deep reinforcement learning techniques to detect and characterize anatomical landmarks in the inner ear, which are vital for guiding CI procedures. These methods enable precise identification of neural structures and head orientation in CT scans, improving the safety and effectiveness of these interventions. We further extended the deep reinforcement approach to detect abnormal anatomies of the inner ear, which was both pioneering work for this anatomical region and also a step forward for anomaly detection in 3D medical images.
Another important contribution of this thesis is the development of the first automated approach for classifying congenital inner ear anomalies. This unsupervised method utilizes latent space representations of 3D point clouds to categorize the different types of congenital malformations, providing latent space representation that is representative of these complex and rare inner ear pathologies. This approach addresses a critical gap in the field, offering support for the diagnosis of these difficult conditions. The thesis also extends its focus to left atrial appendage occlusion procedures, using a graph-based model to analyze 3D cardiac structures. These models are designed to capture key anatomical features that influence procedural outcomes, helping clinicians navigate the complex and variable cardiac anatomy more effectively. We focus on exploiting the explainability features of this method to raise new clinical questions.
Overall, this thesis makes substantial contributions to the field of medical imaging by integrating cutting-edge deep learning techniques with specific clinical practice objectives. The research, not only advances the technical capabilities of medical image analysis but highlights the importance of collaboration between computer scientists and clinical experts to translate these innovations to enhance patient care.