Frequency-Guided U-Net: Leveraging Attention Filter Gates and Fast Fourier Transformation for Enhanced Medical Image Segmentation

TL;DR

This paper introduces GFNet, a novel frequency-guided U-Net model utilizing FFT and attention filter gates to improve medical image segmentation accuracy and efficiency, especially for low-resolution images.

Contribution

GFNet is the first to integrate FFT with attention filter gates in U-Net for enhanced segmentation of low-resolution medical images.

Findings

GFNet achieves a Mean Dice of 0.8366, outperforming baseline models.

Attention Filter Gate improves segmentation accuracy over traditional methods.

FFT integration reduces computational costs and enhances feature extraction.

Abstract

Purpose Medical imaging diagnosis faces challenges, including low-resolution images due to machine artifacts and patient movement. This paper presents the Frequency-Guided U-Net (GFNet), a novel approach for medical image segmentation that addresses challenges associated with low-resolution images and inefficient feature extraction. Approach In response to challenges related to computational cost and complexity in feature extraction, our approach introduces the Attention Filter Gate. Departing from traditional spatial domain learning, our model operates in the frequency domain using FFT. A strategically placed weighted learnable matrix filters feature, reducing computational costs. FFT is integrated between up-sampling and down-sampling, mitigating issues of throughput, latency, FLOP, and enhancing feature extraction. Results Experimental outcomes shed light on model performance. The Attention Filter Gate, a pivotal component of GFNet, achieves competitive segmentation accuracy (Mean Dice: 0.8366, Mean IoU: 0.7962). Comparatively, the Attention Gate model surpasses others, with a Mean Dice of 0.9107 and a Mean IoU of 0.8685. The widely-used U-Net baseline demonstrates satisfactory performance (Mean Dice: 0.8680, Mean IoU: 0.8268). Conclusion his work introduces GFNet as an efficient and accurate method for medical image segmentation. By leveraging the frequency domain and attention filter gates, GFNet addresses key challenges of information loss, computational cost, and feature extraction limitations. This novel approach offers potential advancements for computer-aided diagnosis and other healthcare applications. Keywords: Medical Segmentation, Neural Networks,