Aims of the Project
- Implement a deep learning-based approach to detect and localise craniofacial anomalies in children
- Develop a generative method to simulate a stereotypical paediatric skull from a skull with detected anomalies using a skuull atalas as reference
- Integrate methods into a computer-assisted planning workflow and validate on bespoke patient-specific phantoms
Craniosynostosis is a condition in which one or more skull sutures fuse early. Depending on which skull sutures are fused children have a variety of abnormal skull and face shapes. Reconstruction of a more normal skull shape in children with these abnormalities is a challenging surgical task requiring an in-depth understanding of the 3D geometry of the deformity and the best arrangement of skull pieces to ensure sufficient space for brain development and improving the cosmetic appearance. Current clinical practice requires surgeons to cut the skull as spepcific location then reconfigure the bone fragements in a configuration that achieves the desired skull shape, taking into consideration development of the patient as they age, appropriate pairing with normal skull segments, and the surgical feasibility. Virtual surgical planning methods can help to augment this process by allowing surgeons to view the skull as a 3D rendering and construct appropriate virtual representations of skull pieces to help assess the appropriate surgical approach. Challenges remain in defining exactly which portions of the skull need to be removed and how they can be best rearranged to reconstruct the contour of the skull in an aesthetically pleasing manner whilst providing sufficient space for brain development.
This project focuses on the development of computer algorithms to leverage artificial intelligence techniques, including deep learning and generative models, to help assist surgeons in planning craniofacial reconstructions. The primary focus of this Ph.D. will be in developing deep learning-based methods to localise skull abnormalities and generating appropriate corrective skull configurations for reconstruction. Key challenges include the development of appropriate data representations for surface shape to be used in training a neural network in a low data environment. To overcome the small dataset size, ensure generalisability and avoid overfitting, methods to create simulated skull abnormalities will be developed as a data augmentation strategy to create a larger and more varied training dataset. Such techniques can also be used to create a paired dataset of normal-abnormal skulls in order to learn how to generate normal skull surfaces for a wide range of skull abnormalities. Different methods of encoding surface shapes and geometries will be investigated in order to identify an appropriate data representation to be used for anomaly detection and surface generation. A final component of this project is to work with the team to integrate the developed methods into a computer-assisted planning workflow in order to provide realistic and surgically feasible reconstruction plans. These plans can be validated using 3D printed phantoms of patient-specific anatomy for children with specific skull abnormalities.
The ideal candidate will have a strong computational background with experience in machine learning, deep-learning, or related topics. Interest in data representations for shape and geometry and how to integrate these with deep learning is desired. Close collaboration and interaction with surgeons are expected.