CQformer: Learning Dynamics Across Slices in Medical Image Segmentation.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Deep learning models for 3D medical image segmentation often struggle with continuous variations across 2D slices.
• Existing methods utilize convolutions, Transformers, inter-slice interactions, and time series for segmentation.

Purpose of the Study:

• To propose a novel deep learning architecture, CQformer, for enhanced 3D medical image segmentation.
• To model the continuous variation across 2D slices using ordinary differential equations (ODEs).

Main Methods:

• Developed a cross instance query-guided Transformer architecture (CQformer).
• Incorporated a cross-attention mechanism within an ODE formulation to bridge contiguous 2D slice features.
• Utilized a regression head to optimize the gap between bottleneck and prediction layers.

Main Results:

• CQformer demonstrated superior performance over state-of-the-art methods on 6 out of 7 diverse datasets (CT, MRI; organ, tissue, lesion segmentation).
• Achieved performance improvements ranging from 0.44% to 2.45%.
• Secured the second-highest performance (88.30%) on the BTCV dataset.

Conclusions:

• CQformer effectively leverages features from preceding slices to enhance segmentation of subsequent slices in 3D medical imaging.
• The proposed ODE-based approach offers a promising direction for improving deep learning-based medical image segmentation accuracy.