Dual-perspective decoupling network for kidney tumor segmentation on CT images

Area of Science:

• Medical Imaging
• Computer Vision
• Artificial Intelligence

Background:

• Kidney tumor segmentation faces challenges like unpredictable location, object similarity, and variable boundaries.
• Current methods often fail to address all these challenges effectively.
• Object-agnostic or single-perspective approaches limit segmentation performance.

Purpose of the Study:

• To propose a novel Dual-perspective Decoupling Network (DDNet) for enhanced kidney tumor segmentation.
• To address the limitations of existing methods in handling complex segmentation scenarios.
• To improve the accuracy and precision of kidney tumor delineation in medical images.

Main Methods:

• Developed the Dual-perspective Decoupling Network (DDNet) comprising a Dual-perspective Decoupling Module (DDM) and an Edge Refinement Module (ERM).
• The DDM utilizes a Multi-scale Decoupling Branch (MDB) for body/edge decoupling and an Object Decoupling Branch (ODB) for inter-object decoupling.
• The ERM leverages boundary features to refine low-level encoder features, improving boundary precision.

Main Results:

• DDNet achieved superior performance on the KiTS19, KiTS21, and KAT-Seg datasets.
• Achieved Dice scores of 86.08% (KiTS19), 85.45% (KiTS21), and 88.03% (KAT-Seg).
• Outperformed eleven state-of-the-art methods by at least 1.63%, 0.72%, and 1.72% respectively.

Conclusions:

• DDNet effectively addresses the challenges in kidney tumor segmentation.
• The proposed dual-perspective decoupling and edge refinement strategies significantly improve segmentation accuracy.
• DDNet represents a substantial advancement in automated kidney tumor segmentation technology.