Improved UNet with Attention for Medical Image Segmentation.

Area of Science:

• Medical image analysis
• Deep learning for medical imaging
• Computer-aided diagnostics

Background:

• Medical image segmentation is vital for diagnostics.
• UNet, a CNN-based model, is standard but struggles with long-range dependencies.
• Transformers excel at long-range dependencies but lack local detail, leading to hybrid approaches like TransUNet.

Purpose of the Study:

• To develop an advanced UNet architecture that integrates CNNs and Transformers to overcome limitations in medical image segmentation.
• To enhance feature representation in skip connections and decoders for improved segmentation accuracy.
• To address challenges like blurred boundaries, low contrast, and noise in medical images.

Main Methods:

• Proposed a novel hybrid CNN-Transformer model with architectural improvements.
• Introduced a Three-Level Attention (TLA) module incorporating Attention Gate, channel attention, and spatial normalization.
• Redesigned skip connections inspired by UNet++ and incorporated deep supervision similar to BASNet.

Main Results:

• The proposed model consistently improved UNet's prediction performance on both CT and ultrasound datasets.
• The Three-Level Attention module effectively enriched feature representation.
• Enhanced skip connections and deep supervision reduced the semantic gap and improved segmentation accuracy.

Conclusions:

• The novel hybrid UNet architecture offers superior performance for medical image segmentation compared to standard UNet.
• The integration of advanced attention mechanisms and architectural modifications effectively addresses segmentation challenges.
• The model demonstrates broad applicability across different medical imaging modalities.