MDFormer: a multi-scale dense dilated transformer model for 3D medical image segmentation.

Area of Science:

• Medical Imaging
• Computer Vision
• Artificial Intelligence

Background:

• Existing medical image segmentation models struggle with multi-scale information under resolution constraints.
• There is a need for efficient and accurate segmentation models for clinical diagnosis and treatment.

Purpose of the Study:

• To develop a novel model, the multi-scale dense dilated Transformer (MDFormer), for precise medical image segmentation.
• To overcome limitations in capturing multi-scale information and reduce computational costs without compromising accuracy.

Main Methods:

• Developed the multi-scale dense dilated Transformer (MDFormer) incorporating the multi-scale dense dilated self-attention (MDDSA) module.
• The MDDSA module dynamically adjusts matrix size for spatial downsampling and cross-scale aggregation, reducing computational costs.
• Evaluated the model on ACDC and Synapse datasets.

Main Results:

• Achieved a Dice Similarity Coefficient (DSC) of 92.7% on the ACDC dataset and 86.88% on the Synapse dataset.
• The MDFormer model has 37.7 million parameters and 47.39G FLOPs.
• Demonstrated statistically significant improvements in segmentation performance compared to other models via paired T-tests.

Conclusions:

• The MDFormer model significantly enhances medical image segmentation by effectively leveraging multi-scale information extraction.
• The model shows promising potential for various medical image segmentation tasks.
• Provides a foundation for future advancements in transformer-based medical imaging models.