Deep learning for automated, motion-resolved tumor segmentation in radiotherapy

NU2025-264

INVENTORS

• Mohamed Abazeed*
• Sagnik Sarkar
• Peng Thian Troy Teo

SHORT DESCRIPTION

This invention describes a method for automatically outlining lung tumors for radiotherapy planning. It uses deep learning to precisely identify tumors across breathing cycles, improving accuracy and reducing manual effort.

BACKGROUND

Accurate tumor segmentation is a critical yet challenging step in radiotherapy planning, as manual delineation is labor-intensive, highly variable among clinicians, and prone to errors. This variability, both between and within observers, is exacerbated by respiratory motion, which causes tumors to shift, making consistent and precise target definition difficult and potentially leading to suboptimal dosimetry and patient outcomes. Existing automated and semi-automated approaches, such as atlas-based methods, classical image processing, and many deep learning models, often face significant limitations, including insufficient generalizability across diverse imaging protocols and institutions, inadequate robustness in accounting for complex respiratory motion without relying on potentially error-prone deformable image registration, and a lack of clinically validated prognostic capabilities. These shortcomings underscore a persistent need for more reproducible, accurate, and motion-aware tumor delineation solutions that can seamlessly integrate into clinical workflows.

ABSTRACT

The deep learning system (iSeg) automates motion-resolved tumor segmentation for radiotherapy planning, adaptable across various tumor sites and imaging modalities. It employs a 3D U-Net convolutional neural network with an encoder-decoder architecture, trained on over 1,000 multi-institutional, physician-annotated CT datasets. The system preprocesses volumetric medical images by resampling, normalizing intensity, and stratifying by respiratory phase. It generates voxel-level probability maps for gross tumor volume (GTV) for each phase, then uses ensemble modeling to combine these into an Internal Target Volume (ITV) that accounts for respiratory motion without requiring deformable registration. Post-processing steps include morphological filtering, connected-component analysis, and smoothing. The system achieves a Dice similarity coefficient of 0.70–0.73, comparable to human inter-observer variability, and generates prognostic reports by correlating segmentation discordance with local tumor recurrence.

APPLICATIONS

• Radiotherapy Treatment Planning and Delivery Systems – The invention automats and standardizes motion-resolved tumor segmentation with prognostic insights, directly enhancing the precision and efficiency of radiation treatment planning and delivery workflows.
• Oncology AI and Clinical Decision Support Platforms – The system provides automated, reproducible, and prognostically valuable tumor segmentation, enabling advanced clinical decision support and personalized treatment strategies.
• Biopharmaceutical and Oncology Research – Serves as a critical tool for standardizing tumor assessment in clinical trials and radiomics studies.

ADVANTAGES

• Enhanced accuracy and reproducibility – Improves upon manual segmentation methods, which are labor-intensive and prone to variability.
• Motion-resolved segmentation without deformable registration – Propagates segmentations across all respiratory phases of 4D CT to account for respiratory motion without deformable image registration, a common but error-prone step in existing workflows.
• Reduced radiation exposure to normal tissue – Produces ITVs that are 20-40% smaller than physician-delineated volumes
• Prognostic insights through segmentation discordance – Identifies increased risk of local tumor recurrence, providing prognostic information not typically available in existing tools.
• Seamless integration into clinical workflows – Can be deployed as standalone module or integrated into existing radiotherapy treatment planning systems.

PUBLICATION

Sarkar et al. Deep learning for automated, motion-resolved tumor segmentation in radiotherapy. Npj Precision Oncology, 9: 173. 2025.

IP STATUS

A US provisional patent has been filed.

iSeg neural network architecture and workflow.