RPDNet: A reconstruction-regularized parallel decoders network for rectal tumor and rectum co-segmentation

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Accurate segmentation of rectal tumors and the rectum in MRI is crucial for diagnosis and treatment planning.
• Challenges include variable tumor shapes and unclear boundaries, often exacerbated by information loss in traditional deep learning segmentation models.
• Existing encoder-decoder networks struggle with detailed feature preservation due to downsampling.

Purpose of the Study:

• To propose a novel deep learning network, RPDNet, for accurate co-segmentation of rectal tumors and the rectum.
• To address information loss and improve boundary delineation in rectal MRI segmentation.
• To enhance the precise diagnosis and treatment planning for rectal cancer.

Main Methods:

• Developed a Reconstruction-regularized Parallel Decoder network (RPDNet) with a shared encoder and parallel decoders.
• Incorporated an auxiliary reconstruction branch with consistency loss to preserve anatomical information.
• Introduced a target-adaptive attention module to enhance feature contrast for unclear boundaries.

Main Results:

• RPDNet achieved Dice coefficients of 84.91% for rectal tumor segmentation and 90.36% for rectum segmentation.
• The proposed method outperformed state-of-the-art approaches in segmentation accuracy.
• Demonstrated effective preservation of anatomical structure and boundary information.

Conclusions:

• RPDNet effectively addresses information loss and boundary ambiguity in rectal MRI segmentation.
• The novel architecture and attention mechanism contribute to superior segmentation performance.
• RPDNet shows significant potential for clinical application in rectal cancer diagnosis and treatment.