SUMMARY
Lung cancer is the leading cause of cancer-related deaths in the U.S. Curative radiotherapy + chemotherapy is
the standard of care for patients with inoperable or unresectable disease that has spread beyond the primary
tumor to the lymph nodes. Unfortunately, this treatment approach has a high recurrence of 15%-40% and
advanced treatments including immunotherapy combined with radiation increase toxicity to organs. Spillover
radiation to normal organs at risk (OAR) results from treatment margins to account for uncertainty in localizing
tumors and OARs. Despite being part of standard equipment, information from in-treatment room cone-beam
computed tomography scans (CBCTs) is currently used only in limited ways for patient positioning during
treatment, without simultaneous online localization of the tumor and each OAR. This proposal will use innovative
artificial intelligence (AI) methods, that have been trained from both CT and magnetic resonance imaging (MRI)
studies, to create auto-segmentation tools that can accurately localize the tumor and key OARs online at
treatment setup.
The proposed novel AI methodology is called “Cross-Modality Educed Learning” or CMEDL
(‘c-medal’). The key advantage of CMEDL is that MRI datasets, even from different patients, can be used, to
guide the CT/CBCT network and “learn” to extract features that emphasize the difference between tissue types
and produces accurate segmentations even in areas with little inherent contrast such as the mediastinum.
For
the first time, the clinical utility of what could be called AI-Guided Radiotherapy (AIGRT) segmentation tools will
be systematically studied in relation to their potential impact on treatment margin reduction and normal tissue
toxicity modeling for longitudinally segmented tumor and healthy tissues on CBCTs. Proposed AIGRT tools
would provide increased geometric confidence as well as provide a better basis for an after-delivery estimate of
delivered dose, and treatment toxicity, enabling better risk-benefit assessments for potential treatment
adaptations. Aim 1: Apply CMEDL methodology to develop lung tumor and OAR segmentations on planning
CTs. Aim 2: Extend the CMEDL methodology to longitudinally segment tumors and OARs on weekly CBCTs,
incorporating patient-specific anatomic and shape priors from planning CTs. Aim 3: Determine whether CMEDL
can enable improved (safer) lung cancer radiotherapy dose characteristics by performing automated planning
and delivery simulations, using in-house planning system. Project goal: To develop and rigorously test AIGRT
tools for lung cancer radiotherapy treatments. Potential impact: If successful, these innovative AI tools could be
deployed routinely, enabling (1) smaller margins and less radiotherapy toxicity for patients, including those with
very difficult-to-treat centrally located tumors and (2) providing tools for monitoring the need for plan changes.
These AIGRT tools could potentially be deployed to other disease sites, and once established be made widely
available as a pragmatic, generalizable technology for geometry guidance throughout the radiation treatment.