Slice-prompted HR-CTV interactive segmentation for cervical cancer brachytherapy: A multi-center study

Area of Science:

• Medical Imaging
• Artificial Intelligence in Medicine
• Oncology

Background:

• Manual contouring of high-risk clinical target volume (HR-CTV) in CT-guided cervical cancer brachytherapy is time-consuming and requires expertise.
• Automated methods often struggle with the ambiguous boundaries of HR-CTV.

Purpose of the Study:

• To develop an efficient, interactive segmentation framework that combines deep learning with clinician expertise for HR-CTV delineation.
• To improve the accuracy and efficiency of HR-CTV segmentation in cervical cancer brachytherapy.

Main Methods:

• Proposed a slice-prompted interactive segmentation method (SPSeg) using a 3D U-Net architecture.
• Clinicians provided sparse prompts on key slices to guide full-volume segmentation.
• Investigated two variants: SPSeg-Mono (single encoder) and SPSeg-Dual (dual encoders with deeper feature fusion).

Main Results:

• SPSeg-Dual significantly improved segmentation accuracy, increasing Dice Similarity Coefficient (DSC) from 0.76 to 0.91 and decreasing 95% Hausdorff Distance (HD95) from 11.6 to 3.2 mm with only three prompt slices.
• Reduced contouring time from over 10 minutes to approximately 1.5-1.7 minutes per patient.
• Enhanced inter-observer agreement with DSC increasing from 0.88 to 0.93 and HD95 decreasing from 3.2 to 2.5 mm.

Conclusions:

• The SPSeg method effectively integrates clinical expertise and deep learning for precise and efficient HR-CTV delineation.
• This approach offers a clinically viable solution for improving cervical cancer brachytherapy planning.