FEM-NEURODESIGN - Neuromorphic Approach to the Biomechanical Design of Internal Fixation

Femoral fractures represent one of the leading causes of disability and mortality worldwide, with a steadily increasing incidence among the elderly population due to the rising prevalence of osteoporosis and bone fragility. In Italy, fragility fractures exceed 500,000 cases per year, with a significant proportion involving the proximal femur. The clinical management of these fractures requires prompt and accurate diagnosis, which is essential for defining the most appropriate treatment strategy and for reducing postoperative complications and healthcare costs. However, traditional radiological diagnosis still largely relies on the physician’s subjective experience, leading to inevitable variability and potential diagnostic inconsistencies. In recent years, the evolution of neuromorphics has introduced new perspectives in medical imaging diagnostics, enabling the automation of fracture detection, segmentation, and classification processes. These technologies allow for the recognition of morphological patterns that are not always discernible to the human eye, thereby improving diagnostic accuracy and offering concrete support to clinical decision-making. This thesis aims to develop and train a neural network for the analysis of retrospective femoral X-ray images (RX), with the objective of automatically detecting and classifying proximal femoral fractures. In the first phase, a You Only Look Once (YOLO) model was implemented to automatically detect and classify fractures according to the AO classification system, with a specific focus on types A2 (pertrochanteric) and B2 (transcervical). The YOLO model achieved a Precision of 0.808\% and a Recall of 0.797\% on the test set, demonstrating good detection and classification capability. In the second phase, key morphometric parameters were manually extracted, including the Pauwels angle, neck-shaft angle (NSA), lateral wall thickness (LWT), fracture length, and femoral neck diameter, and integrated with clinical data such as age, osteoporosis, and fracture stability. These features were used to train supervised machine learning models (Random Forest, Support Vector Machine, and k-Nearest Neighbors), optimized through k-fold cross-validation (k = 6), selecting the configuration with the highest balanced accuracy. Among the tested models, SVM achieved the best performance, with a balanced accuracy of 72.82\% in predicting the most appropriate therapeutic treatment. A specific focus was placed on the explainability analysis of the model, which was performed using the SHAP (SHapley Additive exPlanations) framework. This approach demonstrated a high capacity to interpret clinical and morphological parameters in a manner consistent with medical reasoning. The orthopedic specialist involved confirmed that the network successfully replicates the physician’s decision-making logic, thereby validating the robustness and reliability of the system. As a result, this work led to the initiation of a retrospective clinical trial, involving Istituto Ortopedico Rizzoli and Ospedale Maggiore (Bologna), laying the groundwork for its future implementation as a decision-support tool for surgical planning and standardization of therapeutic protocols in orthopedic practice.