APT-Net: Adaptive encoding and parallel decoding transformer for medical image segmentation.

Area of Science:

• Medical Image Analysis
• Artificial Intelligence
• Computer Vision

Background:

• Transformer-based networks face challenges in medical image segmentation due to inadequate position encoding and inefficient contextual information utilization in serial decoders.
• Existing methods struggle with effectively capturing multi-scale features and positional nuances critical for precise segmentation.

Purpose of the Study:

• To propose APT-Net, a novel CNN-transformer hybrid network addressing limitations in medical image segmentation.
• To enhance position encoding and decoding mechanisms for improved segmentation accuracy.
• To achieve state-of-the-art performance on various medical image segmentation tasks.

Main Methods:

• Developed APT-Net, an encoder-decoder architecture integrating CNN and transformer components.
• Introduced an adaptive position encoding module with multi-receptive field fusion for richer positional information.
• Implemented a dual-path parallel decoder to process multiscale features efficiently using basic and guide paths.

Main Results:

• Achieved high Intersection over Union (IoU) scores, including 0.783 on ISIC2017 and 0.851 on Glas datasets.
• Demonstrated state-of-the-art performance on polyp segmentation and the Glas dataset.
• Ablation studies confirmed the effectiveness of the adaptive position encoding and dual-path parallel decoder.

Conclusions:

• APT-Net significantly advances medical image segmentation by effectively encoding positional information and utilizing contextual data.
• The proposed network shows high accuracy and portability across diverse medical imaging datasets.
• APT-Net represents a promising direction for improving transformer-based medical image segmentation models.