PaperlayerView exact revision on arXiv

Source-linked AI summary

Multi-Resolution Feature Fusion U-Net for Magnetic Resonance Imaging Segmentation

Eirini Cholopoulou, Dimitrios E. Diamantis, Dimitris K. Iakovidis

arXiv:2610.00279v1cs.CVcs.LG

TL;DR

MRI segmentation is important for clinical assessment, but manual delineation is costly and anatomical irregularities challenge feature preservation and contextual representation. MRFFU-Net integrates multi-resolution feature fusion with attention and skip connections across a U-Net, and outperforms state-of-the-art methods on CSF and cardiac MRI benchmarks.

Abstract

from arXiv · show

The segmentation of anatomical structures in medical images and particularly in MRI scans, is essential for clinical diagnosis and monitoring disease progression. While Deep Learning (DL) architectures, such as U-Net and its extensions are very effective in medical image segmentation tasks, they often struggle with preserving fine-grained details and global contextual information. This is especially challenging for MRI data segmentation, where anatomical structures are characterized by irregular boundaries and variations in shape, contrast, and scale. To address this challenge, we propose a novel DL architecture for MRI segmentation across different anatomical structures. Specifically, the architecture introduces a module, named Multi-Resolution Feature Fusion (MRFF), that can be easily integrated into any U-Net-like architecture. The MRFF is integrated in all levels of an encode-decoder structure, along with attention mechanisms and skip connections to extract features at multiple resolutions, enabling the model to capture both fine-grained details and global contextual information. We evaluate the MRFFU-Net on two publicly available benchmark MRI datasets of different anatomical targets; one for Cerebrospinal Fluid (CSF) segmentation in spinal MR scans, and one for left atrium cardiac segmentation, from the Medical Segmentation Decathlon (MSD) challenge. Experimental results indicate that MRFFU-Net outperforms state-of-the-art models across multiple evaluation metrics, demonstrating its effectiveness in MRI segmentation.

I. INTRODUCTION

MRI segmentation supports clinical assessment, but manual delineation is costly and CSF’s low-contrast boundaries make segmentation challenging. The paper proposes a multi-resolution U-Net architecture to address gaps in feature fusion across anatomical structures.

I. INTRODUCTION

  • Manual delineation of MRI structures is costly and time-consuming, while CSF’s low-contrast boundaries and similarity to surrounding structures complicate segmentation.The text identifies automated segmentation as a way to assist clinicians.Source: §I · p. 1
  • Reviewed MRI segmentation approaches use mechanisms including skip connections, residual connections, and attention mechanisms, but commonly fuse features through static concatenation or addition.The introduction identifies learning across multiple resolutions as a gap in capturing both fine detail and global context.Source: §I · p. 1
  • Some feature-fusion methods focus on one anatomical structure, while transformer hybrids can add computational overhead, require large training datasets, or fuse features only at selected stages.The introduction notes that selective or absent multi-resolution fusion may hinder capture of both fine and coarse features.Source: §I · pp. 1–2
  • MRFFU-Net integrates multi-resolution feature fusion throughout a U-Net architecture, with attention and skip connections, and is evaluated on CSF and left atrium MRI segmentation.The MRFF modules are intended to capture fine-grained detail and global context.Source: §I · p. 2

II. METHODOLOGY

The methodology presents a U-Net enhanced with MRFF modules, soft attention, and skip connections to support contextual learning and spatial localization.

II. METHODOLOGY

  • The architecture combines a U-Net with MRFF modules to support contextual learning.Source: §II · p. 2
  • Soft attention modules are incorporated alongside the MRFF modules.Source: §II · p. 2
  • Skip connections are combined with the modules to support spatial localization.Source: §II · p. 2

A. Multi-Resolution Feature Fusion Module

The MRFF module extracts features at different receptive fields and fuses them to represent fine-grained and contextual information while retaining the input’s spatial dimensions.

A. Multi-Resolution Feature Fusion Module

B. Model Architecture

The encoder extracts progressively coarser features with MRFF blocks, while an attention-guided decoder restores resolution and uses skip connections and MRFF modules to reconstruct the segmentation.

B. Model Architecture

  • Four encoder MRFF blocks extract small-, medium-, and large-resolution features before downsampling, as convolutional blocks reduce spatial dimensions and increase filter depth.Each encoder block is followed by an MRFF module; the final MRFF output is followed by a convolutional layer with 256 feature-map channels.Source: B. Model Architecture · p. 3
  • The decoder uses transposed convolutions and soft attention to upsample feature maps and select relevant fine-grained details for localization.Source: B. Model Architecture · p. 3
  • At each reconstruction level, encoder residual connections are concatenated with upsampled decoder features, and decoder MRFF modules fuse features across resolutions.Source: B. Model Architecture · p. 3
  • The architecture integrates multi-resolution fusion and soft attention across encoder and decoder paths, while pointwise fusion and residual connections preserve low-level details and support gradient flow.Source: B. Model Architecture · p. 3

C. Optimization Objective

The optimization objective uses Binary Cross-Entropy (BCE) to measure the difference between predicted segmentations and ground-truth masks. Its pixel-level formulation uses each pixel’s ground-truth and predicted labels.

A. Experimental Setup

The evaluation uses two MRI datasets with distinct anatomical targets: spinal-cord CSF and left-atrial segmentation. MRFFU-Net is compared with established segmentation methods using four metrics and dataset-specific train/test procedures.

  • The spinal dataset contributes 2,116 T2-weighted axial slices from 10 patients with expert CSF annotations; the Heart dataset contains 30 MRI scans, with 20 annotated training cases.The spinal scans came from multiple patients and MRI scanners; the Heart dataset was selected for its large anatomical variability.Source: A. Experimental Setup · p. 4
  • The comparison uses MAE, MSE, IoU, and Dice against ArachNet and vanilla U-Net, Attention U-Net, U-Net++, and U-Net 3+.Source: A. Experimental Setup · p. 4
  • CSF experiments use five-fold validation with eight patients for training and two for testing per fold; Heart uses an 80/20 train/test split and 10% validation.Training used Adam, batch size 32, learning rate 0.001, early stopping, 224×224 inputs, and random horizontal and vertical flips.Source: A. Experimental Setup · p. 4

B. Experimental Results

MRFFU-Net achieves the strongest reported results on both CSF and left-atrial MRI segmentation tasks. Qualitative examples show localization across morphological variation, with minor CSF boundary deviations also reported.

  • MRFFU-Net outperforms comparative methods on both tasks, with CSF Dice 0.886, IoU 0.843, MAE 0.011, and MSE 0.005, and Heart Dice 0.879 and IoU 0.816.Source: B. Experimental Results · p. 4, Table II · p. 4, Table I · p. 4
  • For CSF, qualitative results show localization across morphological variation and consistency with ground-truth masks, alongside minor deviations at boundary regions.Source: B. Experimental Results · p. 4
  • On the Heart dataset, qualitative results show consistent left-atrial localization even in high-variability cases, including small and variable structures.Source: B. Experimental Results · p. 4

IV. DISCUSSION AND CONCLUSIONS

MRFFU-Net combines multi-resolution feature fusion with attention and skip residual connections across network levels. The authors report benchmark gains on CSF and cardiac segmentation and identify broader evaluation as future work.

  • The MRFF module fuses features across resolutions to capture fine-grained details and global information for MRI segmentation across anatomical regions.Source: §IV · p. 4
  • MRFF modules at all network levels work alongside soft attention and skip residual connections to refine features and support information flow.The authors state that this design is suited to inconsistent boundaries and intra-class structural variation in MRI scans.Source: §IV · p. 5
  • Future work will evaluate broader medical datasets and domains and explore domain adaptation and semi-supervised learning for limited-annotation settings.Source: §IV · p. 5
Loading 2610.00279v1…