Deep cross-modality (MR-CT) educed distillation learning for cone beam CT lung tumor segmentation

Area of Science:

• Radiotherapy and Medical Imaging
• Artificial Intelligence in Oncology
• Image Segmentation Techniques

Background:

• Cone beam computed tomography (CBCT) is widely available in radiation therapy but lacks reliable segmentation for lung tumors.
• Accurate segmentation is crucial for volumetric response assessment and adaptive radiotherapy.
• Current methods are insufficient for advanced applications beyond basic setup corrections.

Purpose of the Study:

• To develop a novel deep learning method for accurate lung tumor segmentation using CBCT images.
• To improve the utility of CBCT in lung cancer radiotherapy through enhanced segmentation.
• To enable volumetric response assessment and geometry-guided adaptive radiation therapies.

Main Methods:

• Developed a cross-modality educed distillation (CMEDL) deep learning approach.
• Utilized unpaired magnetic resonance imaging (MRI) data to guide CBCT segmentation network training.
• Employed unpaired domain adaptation (UDA) and cross-domain segmentation distillation networks (SDNs) with CycleGAN and contextual losses.
• Compared performance against CBCT-only 2D and 3D networks, and an alternative UDA framework.

Main Results:

• CMEDL significantly improved segmentation accuracy compared to CBCT-only methods (p < 0.001).
• Unet and DenseFCN models using CMEDL achieved higher Surface Dice Similarity Coefficient (SDSC) and lower Hausdorff Distance at 95th percentile (HD95).
• CMEDL outperformed an alternative framework using UDA with an MRI network.

Conclusions:

• The CMEDL approach demonstrates feasibility for accurate lung cancer segmentation from CBCT.
• This method shows potential for advancing lung cancer radiotherapy through improved image analysis.
• Further validation on larger datasets is required for clinical implementation.