Abstract
Medical image segmentation requires a delicate balance between capturing fine-grained local textures and broad global context. While Convolutional Neural Networks (CNNs) excel at the former and Vision Transformers (ViTs) at the latter, both have limitations in efficiency or data requirements. This paper introduces Rolling-Unet, an architecture featuring a novel "Rolling-MLP" module. By cyclically shifting and cropping feature maps, Rolling-Unet captures global dependencies with the computational efficiency of standard $3\times3$ convolutions. Evaluated across four medical imaging datasets, Rolling-Unet achieves state-of-the-art performance, effectively bridging the gap between local precision and global reach.
Introduction
In the domain of medical imaging, understanding the spatial relationship between anatomical structures is as critical as identifying the structures themselves. Traditionally, Convolutional Neural Networks (CNNs) have been the workhorse for segmentation tasks due to their inductive bias for local features. However, their limited receptive fields often fail to capture long-distance dependencies. Conversely, Vision Transformers (ViTs) offer global receptive fields but suffer from quadratic computational complexity ($O(N^2)$) and typically require large-scale pre-training.
Rolling-Unet proposes a return to the Multilayer Perceptron (MLP), reimagined through a "Rolling" operation. This approach allows the model to achieve the global context of a Transformer while maintaining the efficiency and local precision of a CNN.
Methodology
The Rolling-Unet Architecture
The architecture follows a U-Net-like structure, utilizing standard convolutions in the initial layers to capture high-resolution local features. In the deeper layers and the bottleneck—where global context is paramount—the authors introduce the Lo2 (Long-Local) block.
As shown in the architecture diagram, the Lo2 block processes data through two parallel paths:
- DOR-MLP (Long Distance): Captures global dependencies using orthogonal rolling operations.
- DSC (Local Context): Uses Depthwise Separable Convolutions to maintain local spatial awareness.
Crucially, the Lo2 block is designed to be a "drop-in" replacement for standard convolutions, maintaining identical parameter counts and FLOPs while significantly expanding the receptive field.
The Rolling-MLP (R-MLP) Mechanism
The core innovation is the Rolling-MLP. Unlike standard MLPs that flatten spatial dimensions, the R-MLP employs a shifting and cropping strategy. By shifting feature maps by a step $k$ and wrapping the overflow, pixels that were originally far apart are brought into proximity.
By stacking these operations into Orthogonal (OR-MLP) and Double Orthogonal (DOR-MLP) modules, the network systematically captures dependencies in four directions: horizontal (Left-to-Right, Right-to-Left), vertical (Top-to-Bottom, Bottom-to-Top), and diagonals. This directional weight sharing allows the model to learn spatial relationships across the entire image grid efficiently.
Experimental Results
Rolling-Unet was evaluated on four diverse public datasets: BUSI (Breast Ultrasound), GlaS (Gland Segmentation), ISIC 2018 (Skin Lesion), and CHASEDB1 (Retinal Vessel).
Quantitative Performance
Rolling-Unet consistently outperformed state-of-the-art models, including U-Net, UNet++, TransUNet, and UNeXt.
- BUSI Dataset: Rolling-Unet (Large) achieved an IoU of 67.81%, surpassing UCTransNet (67.27%).
- ISIC 2018: At a $512\times512$ resolution, the model achieved an IoU of 84.14%.
Qualitative Insights
Visual analysis confirms that Rolling-Unet is superior at identifying clear boundaries in noisy ultrasound images. Furthermore, it excels at capturing the intricate, thin structures of retinal vessels, where traditional models often produce fragmented or disconnected segmentations.
Discussion and Conclusion
Rolling-Unet demonstrates that MLPs, when augmented with simple spatial rolling operations, can rival the global modeling capabilities of Transformers without the associated computational overhead. The hybrid Lo2 block provides a highly efficient mechanism for multi-scale feature extraction.
Future work will explore extending the "Rolling" concept to 3D medical volumes (CT/MRI) and investigating the architecture's generalizability to other medical imaging tasks such as object detection and classification. While the model shows some sensitivity to overfitting on simpler datasets with larger parameter counts, its overall performance marks a significant step forward in efficient medical image segmentation.
Summary updated