Abstract
Automated 3D medical image segmentation is critical for clinical workflows but often requires significant computational resources. We present MLRU++, a Multiscale Lightweight Residual UNETR++ architecture that optimizes the balance between accuracy and efficiency. MLRU++ introduces a Lightweight Channel and Bottleneck Attention Module (LCBAM) and a Multiscale Bottleneck Block ($M^2B$) to capture multi-resolution features with minimal parameter overhead. Evaluated on benchmarks like Synapse and ACDC, MLRU++ achieves state-of-the-art results, including an 87.57% Dice score on Synapse, while remaining significantly more efficient than traditional hybrid models.
Introduction
The delineation of anatomical structures in volumetric data (CT/MRI) is a cornerstone of modern medicine. While 3D U-Nets and hybrid CNN-Transformer models have advanced the field, their high memory and computational demands often hinder real-time clinical deployment. MLRU++ bridges this gap by redesigning core components to be lightweight without sacrificing the long-range dependency modeling characteristic of modern Transformers. The goal is a model that maintains high precision with a reduced computational footprint.
Methods
MLRU++ is built on a hierarchical encoder-decoder framework. Its primary innovations include:
- Lightweight Channel and Bottleneck Attention Module (LCBAM): Unlike standard attention modules that rely on heavy Multi-Layer Perceptrons (MLPs), LCBAM utilizes $1 \times 1$ and depthwise convolutions. This approach reduces the parameter count for channel attention by 50% while preserving spatial-channel synergy.
- Multiscale Bottleneck Block ($M^2B$): Integrated into the decoder, this block employs parallel depthwise convolutions with varying kernel sizes ($3 \times 3, 5 \times 5, 7 \times 7$). This allows the network to adaptively extract features at multiple scales, which is crucial for medical images containing organs of diverse sizes, such as the gallbladder versus the liver.
- Deep Supervision: The model employs auxiliary losses calculated at multiple decoder stages to improve gradient flow and guide the learning process effectively across different resolutions.
Results and Discussion
MLRU++ was validated across four diverse benchmark datasets: Synapse Multi-organ, BTCV, ACDC (cardiac MRI), and the Decathlon Lung dataset.
On the Synapse multi-organ CT dataset, MLRU++ achieved a state-of-the-art average Dice Similarity Coefficient (DSC) of 87.57% and a Hausdorff Distance (HD95) of 7.53 mm, indicating high boundary precision. Notably, it outperformed its predecessor, UNETR++, by over 5% on this benchmark. On the ACDC cardiac MRI dataset, it reached a 93.00% average Dice score.
The model's efficiency is a primary highlight; it maintains high performance with a significantly lower parameter count and memory footprint compared to heavy-duty models like nnFormer. Quantitative comparisons demonstrate that MLRU++ leads in nearly every category for abdominal organ segmentation while remaining computationally practical.
Conclusion
MLRU++ demonstrates that architectural efficiency and high-precision segmentation are not mutually exclusive. By employing lightweight attention and multiscale processing, the model provides a scalable solution for 3D medical imaging tasks. Future work will explore further optimization for specific edge-computing hardware to enhance its real-time clinical utility.
Summary updated