Resumen de Deep learning-based segmentation methods for computer-assisted fetal surgery
This thesis focuses on the development of deep learning-based image processing techniques for the detection and segmentation of fetal structures in magnetic resonance imaging (MRI) and 3D ultrasound (US) images of singleton and twin pregnancies. Special attention is laid on monochorionic twins affected by the twinto-twin transfusion syndrome (TTTS), which is characterized by the presence of blood vessels located on the surface of the shared placenta. One of the most effective treatments for TTTS is fetoscopic laser photocoagulation, which consists in clogging the blood vessels connecting the twins. Due to the complexity of this intervention, a conscious preoperative planning becomes essential. In particular, selecting the best entry point for the fetoscope, in order to be able to reach all the anastosomes, is key for the intervention.
We propose the first TTTS fetal surgery planning and simulation platform, which incorporates novel algorithms into a flexible C++ and MITK-based application. The software allows full exploration of the intrauterine environment by simulating the fetoscope camera as well as the laser ablation, helping to determine the best entry point for the fetoscope, and training surgeons’ movements and trajectories ahead of operation.
Different approaches are utilized to automatically segment the mother’s soft tissue, uterus, placenta, its peripheral blood vessels, and umbilical cord from multiple (axial, sagittal and coronal) MRI views or a super-resolution reconstruction. In particular, 3D Gabor filters, texture features and Support Vector Machines are employed for placenta segmentation, and curvature-based corner detector for peripheral vessels detection. The mother’s womb is localized and subsequently segmented using reinforcement learning and capsule networks. Finally, the cord is tackled through 3D convolutional networks with long short-term memory (LSTM).
Complementarily, (conditional) generative adversarial networks (GANs) are used for segmentation of fetal structures from (3D) US. A cascade (stack) of GANs is implemented to simultaneously generate synthetic fetal US images from normal and pathological cases, perform multi-class segmentation of the placenta, ecographic shadows, and the peripheral blood vessels, and finally reconstruct (compensate) placenta acoustic shadows. The umbilical cord insertion is also localized from color Doppler US.
Finally, we present a comparative study of deep-learning approaches and Radiomics over the segmentation performance of several fetal and maternal anatomies in both MRI and 3D US. With regard to Radiomics, ninety-four radiomics features are utilized, the optimal ones for each anatomy are identified and fed into a Support Vector Machine with instance balancing.
The work in this thesis has been validated on a large database of clinical cases, with successful results. Although experiments are oriented towards TTTS, results indicate that these methodologies could be extended to other fetal disorders.
