Abstract
Biomedical image segmentation is a critical task often hindered by varying tissue appearances and ambiguous boundaries. While the U-Net architecture is widely adopted, it frequently requires dataset-specific modifications to achieve high performance. This paper introduces ACE-Net, a generalized architecture that augments the standard U-shape with Intensive Context Modeling (ICM) and Multi-Source Aggregation (MSA). By incorporating Atrous Spatial Pyramid Pooling (ASPP) and direct raw-image skip connections, ACE-Net captures multi-scale context and preserves fine-grained spatial details. Experimental results on electron microscopy and retinal imaging datasets demonstrate that ACE-Net achieves state-of-the-art accuracy and high generalizability while maintaining efficient inference times.
Introduction
The U-Net architecture has become the gold standard for medical image segmentation. However, standard implementations often struggle with "intra-class inconsistency"—where the same tissue type appears at different scales or textures—and "edge ambiguity," where boundaries between structures are poorly defined. Many existing solutions address these by creating complex, task-specific cascades that lack generalizability. ACE-Net aims to overcome these hurdles through a robust, generalized architecture that extracts richer features and improves boundary localization without excessive computational overhead.
Methodology
ACE-Net builds upon the traditional encoder-decoder structure, introducing three primary enhancements to the contracting and expansive paths.
Intensive Context Modeling (ICM)
The contracting path is augmented with ICM blocks. These blocks utilize Atrous Spatial Pyramid Pooling (ASPP), which employs convolutional filters with multiple dilation rates. This allows the network to simultaneously capture features across various receptive fields, effectively modeling context at multiple scales.
Multi-Source Aggregation (MSA)
In the expansive path, MSA modules replace standard skip connections. Instead of a single source, MSA aggregates features from four distinct inputs: the preceding decoder block, the corresponding encoder block, dense connections from earlier decoder stages, and the resized raw input image. The inclusion of the raw image provides direct spatial cues that are often lost in deeper convolutional layers, significantly improving boundary definition.
Deep Supervision
To ensure robust learning across all layers, ACE-Net employs deep supervision. Loss is calculated against the ground truth at every block in both the contracting and expansive paths. This strategy serves as a strong regularizer and helps the network learn meaningful intermediate representations.
Experimental Results
The model's performance was validated on two distinct biomedical datasets: the ISBI 2012 EM Segmentation dataset (neuronal structures) and the DRIVE dataset (retinal blood vessels).
ISBI 2012 EM Segmentation
ACE-Net achieved a Vrand score of 0.9797 without post-processing, which increased to 0.9850 with post-processing. This performance represented a significant improvement over standard U-Net baselines, moving up over 30 positions on the competitive leaderboard.
DRIVE Dataset
On the DRIVE dataset, the model demonstrated state-of-the-art Specificity (98.42%) and Accuracy (95.69%), with a competitive AUC of 97.42%. While its sensitivity (77.25%) was slightly lower than some specialized methods, those methods often rely on extensive pre-processing that ACE-Net does not require.
Computational Efficiency
Despite its architectural enhancements, ACE-Net remains computationally efficient. It processes a 512x512 EM image in approximately 0.53 seconds on an NVIDIA TITAN Xp GPU. Its streamlined design allows it to run significantly faster than many other multi-scale segmentation networks.
Conclusion
ACE-Net provides a powerful and generalized framework for biomedical image segmentation. By augmenting the U-Net architecture with ICM and MSA modules, the model effectively addresses multi-scale context and boundary localization. Its consistent performance across different imaging modalities highlights its potential as a robust tool for diverse medical applications. Future work may involve integrating domain-specific pre-processing to further enhance sensitivity in specialized tasks like capillary segmentation.
Summary updated