SRS: Siamese Reconstruction-Segmentation Network Based on Dynamic-Parameter Convolution.

Area of Science:

• Computer Vision
• Medical Imaging
• Artificial Intelligence

Background:

• Dynamic convolution excels in natural image segmentation but struggles with medical image segmentation (MIS) and infrared small target segmentation (IRSTS) due to data limitations and insufficient fitting capacity.
• Existing methods lack the adaptability required for complex segmentation tasks with limited datasets.

Purpose of the Study:

• To introduce dynamic parameter convolution (DPConv) for improved medical and infrared image segmentation.
• To enhance the fitting capacity of dynamic convolutions for specialized segmentation applications.

Main Methods:

• Developed dynamic parameter convolution (DPConv) that leverages deep encoder features from reconstruction tasks to generate adaptive kernels.
• Proposed a siamese reconstruction-segmentation (SRS) network integrating DPConv for enhanced segmentation.
• Conducted experiments on seven diverse datasets (five medical, two infrared).

Main Results:

• DPConv demonstrated superior fitting capacity compared to standard dynamic convolutions.
• The SRS network achieved state-of-the-art performance on multiple medical and infrared segmentation benchmarks.
• Zero-shot segmentation results highlighted the generalization capabilities of DPConv across unseen modalities.

Conclusions:

• DPConv significantly boosts segmentation performance, particularly in data-limited scenarios like MIS and IRSTS.
• The proposed SRS network offers a robust solution for challenging image segmentation tasks.
• DPConv shows promise for broad applications requiring adaptive feature representation.