Optimization of Radiation Therapy Re-Planning with Deep Learning Image Processing
Brief Description of Project:
Intensity modulated radiation therapy (IMRT) is standard of care for curative intent treatment of locally advanced lung and other thoracic malignancies and has been repeatedly shown to increase treatment efficacy while decreasing both short and long term treatment-related toxicity. IMRT ensures that a critical radiation dose reaches targeted regions while maintaining sub-critical dose for vulnerable structures (i.e., “organs at risk”); hence, patients undergoing IMRT suffer from lower morbidity than 3-D conformal radiotherapy (3D CRT) particularly when treating to higher radiation doses. These improved outcomes are due, in part, to the high precision and steep dosimetric gradients achieved with IMRT. A major concern with IMRT is the failure to select and delineate the organs at risk and targets accurately. This failure can occur during initial treatment planning but also indirectly when treatment associated-anatomical change (e.g., tumor shrinkage or patient weight loss) results in incongruence between intended and actual dose delivered to tumor and organs at risk. Therefore, mapping patient-specific anatomy is essential for clinical care because precise localization of anatomical structures provides IMRT dose planning software the spatial information necessary to maintain efficacy while minimizing impact to vulnerable tissues.
Automated methods to identify thoracic regions have improved IMRT outcomes and are now being commercialized for routine clinical practice. It is important to make a distinction between labeling and contouring. Labeling involves assignment of voxel-wise categorical values to an image such that the spatial extent of each structure is marked, while contouring is an IMRT specific process that defines the spatial extent of dose limits. Contours are based on anatomical labels but are adjusted using expert knowledge of therapeutic margins and patient specific disease processes. In the current IMRT workflow, a radiation oncologist edits automatically generated labels to form contours to take into account patient-specific safety margins and IMRT targets (i.e., structures with intended high dose). Current practice for thoracic IMRT is for a radiation oncologist to manually draw contours interactively. Even with modern computer assistance, this is a time consuming process. Therefore, it is typically done once before the start of the course of treatment.
We propose to apply recent advances in robust / deep learning medical image processing to improve automatic labeling of vulnerable thoracic structures for intensity modulated radiation therapy (IMRT) dose planning in the context of adaptive re-planning. We will create these tools to better model patient anatomy, detect when re-planning is necessary, and enhance precision interventions during re-planning.
In this project, we will revisit four successful clinical studies (BRAINICU and MINDUSA) for which CT imaging was not a primary study consideration. We will retrieve imaging data for these participants and perform data wrangling to harmonize the data into a form suitable for machine learning. We will then applied state of the art deep learning networks to detect associations between baseline imaging data and primary study outcomes.
Strong programming experience. Experience with medical imaging and deep learning is preferred.
Nature of Supervision:
Close collaboration with graduate students.
Weekly meetings with faculty.
Weekly group meetings.
A Brief Research Plan (period is for 10 weeks):
We will curate de-identified medical imaging data, create deep learning models, and characterize model performance.
Number of Open Slots: 1
Name: Bennett Landman
Title: Associate Professor
Department: Electrical Engineering
Campus Address: FGH 372
Phone: (615) 322-2338