HNOSeg-XS: Extremely Small Hartley Neural Operator for Efficient and Resolution-Robust 3D Image Segmentation.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computational Science

Background:

• Convolutional Neural Networks (CNNs) and transformers dominate medical image segmentation but face limitations like high computational cost, memory demands, and suboptimal performance at higher resolutions due to discrete training methods.
• CNNs capture long-range correlations through sequential layers, while transformers use multi-head attention, both leading to computational challenges and potential input size reduction, which can compromise segmentation quality.

Purpose of the Study:

• To introduce HNOSeg-XS, a novel, resolution-robust architecture for medical image segmentation.
• To address the limitations of existing CNN and transformer models regarding computational efficiency, memory usage, and performance consistency across different resolutions.

Main Methods:

• The HNOSeg-XS architecture models image segmentation using learnable partial differential equations implemented via a Fourier neural operator, leveraging its zero-shot super-resolution property.
• The approach replaces the Fourier transform with the Hartley transform, reformulating the segmentation problem in the frequency domain for enhanced resolution robustness.

Main Results:

• HNOSeg-XS demonstrates superior resolution robustness with fewer than 34.7k model parameters.
• The model achieved the best overall inference time (< 0.24 s) and memory efficiency (< 1.8 GiB) on BraTS'23, KiTS'23, and MVSeg'23 datasets using a Tesla V100 GPU.
• Outperformed tested CNN and transformer models in terms of speed, memory, and parameter efficiency.

Conclusions:

• HNOSeg-XS offers a fast, memory-efficient, and extremely parameter-efficient solution for medical image segmentation.
• The proposed architecture overcomes the resolution-dependent performance issues of traditional models, providing robust segmentation across various resolutions.
• HNOSeg-XS presents a promising alternative to current dominant models for medical image segmentation tasks.