Volumetric Image Segmentation for Planning of Therapies: Application to dental implants and muscle tumors

Image segmentation, partitioning an image to consistent, meaningful segments, is a requirement for systematic analysis of its contents. Segmentation is used in medical diagnostics and as presented in this work, in treatment planning and therapy assessment. This work presents three robust and fast methods for two applications. The first two methods were designed facial bones to speed up dental implant planning workflows and the third for muscle tumors (uterine fibroids) to automate the mid- and post-treatment analysis of the results of ultrasound therapy.

Both facial bone structures and muscle tumors can take individual, even unpredictable shapes. The used volumetric (three-dimensional) imaging modalities may suffer from distortions and other types of losses of quality because of the constraints set by feasible exposure or available scanning time. A valid, clinical-grade segmentation method should solve the problem fast to minimize wait times in the therapy planning workflow or almost real time when used to update the plan during the therapy.

To meet these needs we first developed a method that is capable segmenting mandibles from Narrow-Beam Volumetric Tomography images. It works by deforming a pre-constructed surface model around the mandibular bone. Our requirements were later upgraded to include all visible facial bones in Cone-Beam Computed Tomography images. For this revised goal we developed a novel datadriven method that reconstructs facial bone surfaces from continuous patches and bridges over holes due to missing teeth or image distortions. When our target shifted from the mandibular bone to the muscle tumor segmentation from Magnetic Resonance images, we were able to carry over the core properties of the algorithm to the new problem successfully.

We verified the robustness of both facial bone and tumor segmentation with independent training and validation sets and found their accuracy to match other published work. The requirement for a very tight computing budget was reached with as fast as under a minute processing time per image volume.