Background: In 2021, the WHO reported that caesarean sections accounted for over 21% of births and could reach nearly one-third by 2030. Although essential when medically indicated, caesarean sections remain surgical procedures with potential long-term risks. Retrospective studies suggest that about 18% are performed due to dystocia, a complication characterized by prolonged labor and often linked to fetal malposition or cephalopelvic disproportion. Research has also shown that maternal lower-limb positioning can influence birth outcomes in cases of dystocia. However, the use of this non-invasive technique remains largely empirical and would benefit from accurate 3D measurement and visualization of the maternal pelvis and fetal skull to enable more standardized and quantitative assessment. Despite this need, related research remains scarce, and available imaging data on near-term pregnancies are limited. In this context, our research team has initiated the development of a system for 3D visualization of the pelvis and fetal head to support dystocic childbirth monitoring.

Thesis Objective: The present thesis, as part of this broader effort, focuses primarily on the development of new methods for the automatic reconstruction of a patient-specific 3D fetal head model during childbirth.

Contributions: We developped a 3D-tracked probe ultrasound system that combines optical tracking and standard 2D ultrasound imaging. Our first contribution is a mathematical analysis and practical evaluation of a novel calibration problem, based on a parallel-wires phantom, aimed at determining the transformation between an optical marker and the ultrasound image. Our second contribution introduces a new methodology for fetal head model reconstruction from ultrasound measurements and its evaluation. It comprises the design of a novel measurement protocol and the acquisition of paired ultrasound and MRI data from pregnant women near term. Furthermore, we developed a statistical surface model using spherical harmonics to interpolate point clouds from fetal skull contours. Our third contribution is the integration of our 3D reconstruction method with deep learning models trained to segment bone contours and to automatically predict fetal head position and orientation on 2D ultrasound images of a target anatomical region.

Results: The 3D localization of ultrasound image pixels using our measurement system demonstrated an average accuracy of 2.67 mm RMSE across the entire image plane. Surface model reconstruction from 3D-tracked ultrasound images achieved clinically consistent performance on near-term pregnant volunteers, achieving an average per-case RMSE of 2.1 mm and an average per-case Hausdorff distance of 4.6 mm for vertex placement. The automated processing pipeline maintained reconstruction quality comparable to manual annotation, showing only minimal degradation in Hausdorff distance, with per-case deviations inferior to 1 mm.