Use of Artificial Intelligence systems to improve planning and navigation tools for orthopedic surgery

Orthopedic surgical planning depends on the ability to extract, reconstruct, and reason over visual representations of patient anatomy. This thesis explores how modern deep learning techniques can advance each stage of this process within an industrial context, contributing to the development and extension of AI-driven tools for computer-aided surgical planning. The work spans three interconnected research directions, consolidated into four scientific contributions. The first direction investigates how architectural choices in CNN design influence segmentation quality in the medical domain, with particular attention to the challenge of accurately delineating fine boundaries and small anatomical structures. Two novel architectures are proposed and validated across multiple imaging modalities and anatomical domains, demonstrating that targeted architectural innovations can improve segmentation accuracy while maintaining computational efficiency suitable for clinical deployment. The second direction explores the potential of generative models to recover anatomical information that is no longer directly observable due to pathology. Focusing on glenoid reconstruction in cases of advanced glenohumeral arthritis, deep learning inpainting is investigated as a patient-specific alternative to population-based statistical methods, establishing its feasibility and quantitative advantages for computer-aided surgical planning. The third direction addresses the integration of visual intelligence into the broader surgical workflow, exploring how heterogeneous information sources — structured knowledge, technical documentation, and product imagery — can be fused within a retrieval-augmented generation framework to support prosthesis classification and evidence-based decision-making. Together, these contributions demonstrate how visual representation — through segmentation, reconstruction, and retrieval — can be leveraged to build more capable and clinically grounded AI tools for orthopedic surgical planning.