A proposed framework for consensus-based lung tumour volume auto-segmentation in 4D computed tomography imaging.

Area of Science:

• Medical Imaging
• Radiotherapy Oncology
• Computational Anatomy

Background:

• Accurate lung tumor delineation on 4D-CT is crucial for radiotherapy planning.
• Manual segmentation is time-consuming and subject to inter-observer variability.
• Automated methods are needed to improve efficiency and consistency.

Purpose of the Study:

• To propose and validate an automated framework for lung tumor volume auto-segmentation using 4D-CT.
• To leverage multi-physician ground truth (GT) estimates for robust auto-segmentation.
• To expedite the delineation of gross tumor volumes (GTV) and internal gross target volumes (IGTV).

Main Methods:

• Constructed multi-expert GT using the STAPLE algorithm from 6 physicians' contours on 4D-CT data of 10 NSCLC patients.
• Employed a deformable model to propagate GT across respiratory phases for auto-segmentation.
• Assessed accuracy using graph cuts for 3D reconstruction and point-set registration.

Main Results:

• STAPLE-based auto-segmented GTV achieved 81.51% to 97.27% volumetric overlap with GT.
• Auto-segmented IGTV demonstrated improved accuracy (90.87% to 98.57% overlap) with reduced variance.
• Segmentation accuracy was independent of the initial propagation phase selection.

Conclusions:

• The proposed framework provides accurate and reliable auto-segmentation of lung tumor volumes (GTV and IGTV) on 4D-CT.
• The method effectively mitigates inter-/intra-observer variability compared to manual segmentation.
• Further development is needed to simplify the workflow for clinical practice.