Are Vision-xLSTM-embedded U-Nets better at segmenting medical images?

Area of Science:

• Medical Imaging
• Computer Vision
• Artificial Intelligence

Background:

• Medical image segmentation traditionally relied on Convolutional Neural Networks (CNNs).
• Emerging hybrid models combine CNNs with Vision Transformers (ViTs), but ViTs present high computational and storage demands.
• There is a need for efficient, high-performing segmentation architectures deployable on resource-constrained systems.

Purpose of the Study:

• To propose Vision Extended Long Short-Term Memory (Vision-xLSTM) as an efficient backbone for medical image segmentation.
• To introduce and evaluate a novel hybrid architecture, U-VixLSTM, integrating CNNs with Vision-xLSTM.
• To demonstrate that U-VixLSTM achieves superior performance with reduced computational costs compared to existing state-of-the-art methods.

Main Methods:

• Developed the U-VixLSTM architecture, integrating CNNs with Vision-xLSTM blocks.
• Vision-xLSTM blocks process CNN feature map patches to capture temporal and global relationships.
• A convolutional feature reconstruction path upsamples Vision-xLSTM outputs for segmentation.

Main Results:

• U-VixLSTM demonstrated superior performance on the Synapse, ISIC, and ACDC medical image segmentation datasets.
• The proposed model achieved effective segmentation results without the substantial computational burden associated with Vision Transformers (ViTs).
• The architecture proved computationally efficient, suitable for deployment in resource-limited environments.

Conclusions:

• U-VixLSTM presents a promising alternative to ViTs for medical image segmentation.
• The model offers a balance of high performance and computational efficiency.
• This approach facilitates deployment in healthcare settings for faster, resource-efficient medical image analysis and diagnosis.