AMTA-Net

AMTA-Net performs segmentation of the prostate bed in computed tomography (CT) images to support accurate post-prostatectomy radiotherapy planning and reduce radiation exposure to organs-at-risk such as the bladder and rectum.

Key Features:

• Concurrent Segmentation: Performs simultaneous segmentation of the prostate bed and adjacent organs-at-risk (OARs) to leverage anatomical interdependencies for improved accuracy.
• Attention Mechanism: Incorporates cascaded attention modules that hierarchically transfer and refine features from segmented OARs to enhance prostate bed delineation and adaptively learn discriminative representations at challenging boundaries.
• U-Net Backbone: Uses a U-Net architecture as the low-level segmentation backbone for initial OAR delineation in CT images.
• Expert-Mimicking Strategy: Emulates radiologists' strategies for delineating non-contrast, variable-boundary targets by transferring context from clearly defined OARs to the virtual prostate bed target.

Scientific Applications:

• Post-prostatectomy radiotherapy planning: Enables precise prostate bed delineation to inform target volume definition in postoperative radiotherapy.
• Organs-at-risk sparing: Improves discrimination of prostate bed boundaries to reduce radiation exposure to OARs such as the bladder and rectum.
• Evaluation and benchmarking: Has been evaluated on a clinical dataset of 186 CT images and reported performance exceeding atlas-based segmentation methods and other deep learning approaches for clinically critical prostate bed boundaries.

Methodology:

The method employs a two-tiered computational approach: a low-level U-Net backbone for segmentation of neighboring OARs, and a high-level attention sub-network using cascaded attention modules to transfer and refine features from segmented OARs for prostate bed segmentation.