Lightweight Multi-Stage Aggregation Transformer for robust medical image segmentation.

Area of Science:

• Medical Imaging
• Computer Vision
• Artificial Intelligence

Background:

• Accurate medical image segmentation requires capturing multi-scale features.
• Hybrid networks combining Convolutional Neural Networks (CNNs) and Transformers offer complementary benefits.
• Existing hybrid models face challenges with high computational costs or suboptimal performance in lightweight versions.

Purpose of the Study:

• To propose a robust, lightweight, multi-stage hybrid architecture for efficient and accurate medical image segmentation.
• To address the limitations of current methods in balancing network complexity, computational cost, and segmentation performance.
• To enhance the extraction of multi-scale features through progressive aggregations for handling complex image variations.

Main Methods:

• Introduced the Multi-stage Aggregation Transformer version 2 (MA-TransformerV2), a novel lightweight multi-stage hybrid architecture.
• Employed parallel lightweight Transformer (Trans) blocks and CNN blocks in a dual-branch encoder for each stage.
• Incorporated a vector quantization block at the bottleneck to discretize features and reduce redundancy.

Main Results:

• MA-TransformerV2 demonstrated superior segmentation accuracy and model capacity compared to state-of-the-art methods.
• The proposed method achieved high performance while maintaining a low computational cost.
• Experimental results on public datasets validated the effectiveness and efficiency of the MA-TransformerV2 architecture.

Conclusions:

• The MA-TransformerV2 effectively extracts multi-scale features using a hybrid CNN-Transformer approach with progressive aggregations.
• The architecture achieves a favorable balance between segmentation accuracy, computational efficiency, and model interpretability.
• This lightweight hybrid model represents a significant advancement in medical image segmentation technology.