DAUNet: A Lightweight UNet Variant with Deformable Convolutions and Parameter-Free Attention for Medical Image Segmentation

arXiv ID: 2512.07051v1

Published: 2025-12-07

Authors: Adnan Munir, Shujaat Khan

Categories: cs.CV, cs.AI, cs.LG

Relevance Score: 0.98 / 1.00

View on arXiv Download PDF

Summary

DAUNet is a novel, lightweight UNet variant designed for medical image segmentation that integrates Deformable V2 Convolutions and Parameter-Free Attention (SimAM) to enhance spatial adaptability and context-aware feature fusion without increasing model complexity. It significantly outperforms state-of-the-art models on challenging medical datasets (fetal ultrasound, CT pulmonary embolism) in segmentation accuracy and efficiency, proving suitable for resource-constrained clinical deployment. The model demonstrates robustness to missing context and low-contrast regions.

Medical Relevance

Accurate medical image segmentation is crucial for automating diagnostic processes, precise treatment planning, and monitoring disease progression. DAUNet's lightweight design and enhanced segmentation performance on critical tasks like fetal health assessment and pulmonary embolism detection can lead to faster, more reliable, and accessible clinical decision-making, particularly in resource-limited environments.

AI Health Application

The AI application is medical image segmentation, which is critical for computer-aided diagnosis and treatment planning. Specifically, DAUNet aims to improve the automated analysis of ultrasound images for fetal head and pubic symphysis, and CT scans for pulmonary embolism detection, with a focus on efficient and robust deployment in clinical environments for real-time applications.

Key Points

**Novel Lightweight Architecture:** DAUNet is a UNet variant specifically engineered to be lightweight while improving medical image segmentation performance.
**Deformable Convolutions:** It incorporates Deformable V2 Convolutions within its bottleneck to dynamically adapt to geometric variations in anatomical structures, enhancing spatial flexibility.
**Parameter-Free Attention (SimAM):** The decoder and skip pathways are augmented with SimAM attention modules for saliency-aware feature refinement, crucially without adding to the model's parameter count or complexity.
**Extensive Evaluation:** The model was rigorously tested on two distinct and challenging medical datasets: FH-PS-AoP (fetal head and pubic symphysis ultrasound) and FUMPE (CT-based pulmonary embolism detection).
**State-of-the-Art Performance:** DAUNet consistently achieves superior performance metrics (Dice score, HD95, ASD) compared to existing state-of-the-art models on the evaluation datasets.
**Parameter Efficiency:** Despite its enhanced performance, DAUNet maintains excellent parameter efficiency, aligning with its lightweight design goal.
**Clinical Robustness:** The model demonstrates strong robustness in segmenting medical images with challenging characteristics such as missing context and low-contrast regions, as well as suitability for real-time and resource-constrained clinical settings.

Methodology

DAUNet is a UNet-based encoder-decoder architecture. Its core methodological advancements involve: 1) integrating Deformable V2 Convolutions in the bottleneck to dynamically adjust the receptive field for better handling of geometric variations; and 2) embedding Parameter-Free Attention (SimAM) modules in the decoder and skip pathways to perform saliency-aware feature refinement, enhancing context awareness without increasing model parameters. The model's performance was evaluated using Dice score, HD95, and ASD on FH-PS-AoP and FUMPE datasets, with comparisons against state-of-the-art models and ablation studies to confirm individual component contributions.

Key Findings

DAUNet significantly outperforms state-of-the-art medical image segmentation models in Dice score, HD95, and ASD on both challenging fetal ultrasound and CT pulmonary embolism datasets. This superior performance is achieved while maintaining exceptional parameter efficiency, establishing DAUNet as a lightweight and highly accurate solution. Ablation studies confirmed that both Deformable V2 Convolutions and SimAM attention modules contribute positively to the model's enhanced segmentation capabilities and its robustness to challenging image conditions.

Clinical Impact

DAUNet's lightweight nature and high accuracy have the potential to significantly enhance clinical workflows. It can enable the deployment of advanced segmentation capabilities in resource-constrained environments (e.g., portable ultrasound devices, remote clinics) and facilitate real-time diagnostics. This translates to faster and more accurate diagnoses, such as early detection of pulmonary embolisms or precise fetal measurements, improving patient outcomes and potentially reducing the burden on clinical staff.

Limitations

The abstract does not explicitly state any limitations of the DAUNet model. Instead, it emphasizes its robustness to challenging conditions like missing context and low-contrast regions, presenting these as problems it effectively addresses.

Future Directions

The abstract does not explicitly mention future research directions for DAUNet.

Medical Domains

Fetal Ultrasound Imaging Radiology (CT imaging) Pulmonary Embolism Detection Diagnostic Imaging Obstetrics

Keywords

Medical Image Segmentation UNet Variant Deformable Convolutions Parameter-Free Attention SimAM Lightweight Model Real-time Segmentation Deep Learning

Abstract

Medical image segmentation plays a pivotal role in automated diagnostic and treatment planning systems. In this work, we present DAUNet, a novel lightweight UNet variant that integrates Deformable V2 Convolutions and Parameter-Free Attention (SimAM) to improve spatial adaptability and context-aware feature fusion without increasing model complexity. DAUNet's bottleneck employs dynamic deformable kernels to handle geometric variations, while the decoder and skip pathways are enhanced using SimAM attention modules for saliency-aware refinement. Extensive evaluations on two challenging datasets, FH-PS-AoP (fetal head and pubic symphysis ultrasound) and FUMPE (CT-based pulmonary embolism detection), demonstrate that DAUNet outperforms state-of-the-art models in Dice score, HD95, and ASD, while maintaining superior parameter efficiency. Ablation studies highlight the individual contributions of deformable convolutions and SimAM attention. DAUNet's robustness to missing context and low-contrast regions establishes its suitability for deployment in real-time and resource-constrained clinical environments.

Comments

11 pages, 7 figures