James Balter Laboratory

Our lab investigates the development of patient models that support both the planning of radiation therapy as well as image-guided patient positioning. Patient models are derived from a combination of the intensity patterns of different tissues across different MRI contrasts, as well as through the integration of prior knowledge of anatomic shapes (e.g. pelvic bones) within a population of patients. Synthetic CT models of the head have been developed and clinical implementation is expanding from initial support of whole-brain radiotherapy to more focal treatments of glioblastomas and stereotactic treatments of multiple metastases.

Example of synthetic CT of the head

Comparison of treatment plans optimized using CT and synthetic CT (MRCT) image volumes for glioblastoma multiforme (GBM)

Alignment of MRCT to Cone Beam CT for positioning of a patient undergoing treatment for brain metastases

In the pelvis, statistical shape models have been used to improve the accuracy and reduce the scanning complexity of synthetic CT modeling. Principal Component Analysis of a statistical bone shape model extracted from deformable modeling of pelvic volumes across a population of previous patients yields a simple description of a bone mask that separates bone from air in the pelvis. This methodology has been demonstrated to provide sufficient accuracy for IMRT planning in the female pelvis, and is currently being translated to a robust software tool for clinical implementation.

Pelvic imaging: MRCT vs. CT

U-net architecture for synthetic CT generation. Each block represents a convolution operation, followed by batch normalization and Leaky ReLU. Last convolution operation converts 64 dimensional channel to 3 channel synthetic CT.

Schematic flow chart of the algorithm for MRI-based synthetic CT (sCT) generation. The training stage, which is consisted of four generators and two discriminators, is shown on the left. Each generator includes several dense blocks. The synthesizing stage is shown on the right, in which a new MR image is fed into the well-trained model to produce the sCT.

MR imaging is typically slow and subject to artifacts from several factors that limit its utility for supporting a number of radiation therapy tasks, including target definition and understanding and reacting to motion.  Fortunately, sufficient sparsity exists to potentially support these tasks with more efficient sampling and/or intelligent reconstruction techniques. Our lab is currently investigating sparsity of biological mapping of diffusion, morphological mapping of anatomic structure, and mapping hierarchical motion states of patients.

Clusters of different tissue types classified to establish prior characteristic diffusion attenuation curves.

MR imaging is typically slow and subject to artifacts from several factors that limit its utility for supporting a number of radiation therapy tasks, including target definition and understanding and reacting to motion.  Fortunately, sufficient sparsity exists to potentially support these tasks with more efficient sampling and/or intelligent reconstruction techniques. Our lab is currently investigating sparsity of biological mapping of diffusion, morphological mapping of anatomic structure, and mapping hierarchical motion states of patients.

Comparison of conventional image reconstruction (top row) with Compressed Sensing reconstruction based on analysis of Principal Components from a population of prior patients.

MR imaging is typically slow and subject to artifacts from several factors that limit its utility for supporting a number of radiation therapy tasks, including target definition and understanding and reacting to motion.  Fortunately, sufficient sparsity exists to potentially support these tasks with more efficient sampling and/or intelligent reconstruction techniques. Our lab is currently investigating sparsity of biological mapping of diffusion, morphological mapping of anatomic structure, and mapping hierarchical motion states of patients.