Multi-Aperture Transformers for 3D (MAT3D) Segmentation of Clinical and Microscopic Images.

Area of Science:

• Biomedical Imaging
• Medical Image Analysis
• Computational Biology

Background:

• Accurate 3D segmentation of biological structures is essential for understanding cellular and organ functions in biomedical imaging.
• Existing methods often struggle with complex structures, necessitating improved segmentation techniques.

Purpose of the Study:

• To introduce a novel 3D segmentation method, MAT3D (Multi-Aperture representation with Transformers for 3D), for biological images.
• To evaluate MAT3D's performance on clinical and organoid datasets, comparing it to existing literature.

Main Methods:

• MAT3D integrates Multi-Aperture representation with Transformer networks and Convolutional Neural Networks (CNNs).
• This hybrid approach combines global context awareness with local feature extraction for precise delineation of biological structures.
• The method was evaluated on the ACDC and Synapse multi-organ segmentation datasets and an organoid dataset with four breast cancer subtypes.

Main Results:

• On clinical datasets (ACDC, Synapse), MAT3D achieved Dice scores of 93.34±0.05 and 89.73±0.04, outperforming existing methods with fewer parameters.
• For organoid segmentation, MAT3D obtained a Dice score of 95.12±0.02 and a PQ score of 97.01±0.01.
• The method significantly reduced parameters to 40 million, enhancing efficiency.

Conclusions:

• MAT3D offers a powerful and efficient solution for 3D segmentation of complex biological structures in biomedical imaging.
• The method demonstrates superior performance on diverse datasets, including clinical and organoid imaging.
• The availability of the code facilitates further research and application in the field.