Iterative Label Denoising Network: Segmenting Male Pelvic Organs in CT From 3D Bounding Box Annotations

Area of Science:

• Medical Imaging
• Radiotherapy
• Computer Vision

Background:

• Accurate segmentation of prostate and surrounding organs (bladder, rectum) in CT images is crucial for effective prostate cancer radiotherapy.
• Current leading methods, Fully Convolutional Networks (FCNs), require large datasets with detailed voxel-wise annotations, which are time-consuming and costly to obtain.
• This annotation bottleneck hinders the development of accurate segmentation models for clinical use.

Purpose of the Study:

• To develop a novel weakly supervised segmentation approach for organs in CT images.
• To reduce the reliance on extensive voxel-wise annotations by utilizing simpler 3D bounding box annotations.
• To improve the efficiency and applicability of automated segmentation in clinical radiotherapy.

Main Methods:

• Proposed a weakly supervised segmentation method using only 3D bounding box annotations.
• Introduced a label denoising module integrated into an iterative training scheme for a label denoising network (LDnet).
• The LDnet iteratively refines segmentation by identifying and preserving reliable training voxels, discarding noisy labels within bounding boxes.

Main Results:

• Achieved high Dice Similarity Coefficients (DSCs): ~94% for prostate, ~91% for bladder, and ~86% for rectum.
• Performance is comparable to fully supervised methods trained on high-quality voxel-wise annotations.
• Outperformed several state-of-the-art segmentation approaches.

Conclusions:

• This work presents the first voxel-wise segmentation in CT images using only 3D bounding box annotations.
• The proposed method significantly reduces manual labeling effort, addressing a key challenge in clinical applications.
• The approach demonstrates the potential to streamline radiotherapy planning and improve patient outcomes.