Introduction
In the early stages of the COVID-19 pandemic, rapid and accurate diagnosis was critical for patient management and containment. While RT-PCR tests served as the gold standard, their sensitivity was often limited, and results could be slow. Chest CT imaging emerged as a powerful alternative, revealing characteristic patterns such as "ground-glass opacities" (GGO) that are indicative of infection.
However, manually delineating these infection regions is an arduous task for radiologists due to the high volume of patients and the visual complexity of the images. Lesions often exhibit low contrast and vary significantly in shape and position across different patients. Standard deep learning architectures, such as the basic U-Net, frequently struggle to capture these subtle, multi-class features effectively. This study introduces a Residual Attention U-Net designed to automatically segment and quantify different types of COVID-19 lung infections from chest CT scans with high precision.
Methodology
The proposed architecture enhances the standard U-Net framework by incorporating two major advancements in deep learning: Aggregated Residual Transformations and Locality Sensitive Hashing (LSH) attention.
Architectural Innovations
- Aggregated Residual Transformations (ResNeXt): Instead of standard convolutional layers in the encoder, the model utilizes ResNeXt blocks. These blocks leverage the concept of "cardinality"—splitting the input into several parallel, lower-dimensional paths before merging them. This allows the network to learn more expressive features without the optimization difficulties associated with making a network excessively deep or wide.
- Locality Sensitive Hashing (LSH) Attention: Standard "soft" attention mechanisms are computationally expensive, scaling quadratically with the number of pixels. To make attention feasible for high-resolution CT slices, the authors implemented LSH attention. This technique uses hashing to group similar feature vectors into "buckets," allowing the model to calculate attention only within those buckets, which drastically reduces memory usage and computational overhead.
Multi-Class Segmentation Strategy
Unlike binary segmentation models that only distinguish between infection and healthy tissue, this model employs a multi-class cross-entropy loss. This strategy allows the simultaneous identification of three distinct radiographic features: ground-glass opacification (GGO), consolidations, and pleural effusions.
Experimental Setup
Dataset
The study utilized a public dataset provided by the Italian Society of Medical and Interventional Radiology (SIRM), consisting of 110 axial CT images from 60 patients. Ground-truth masks were annotated by expert radiologists. Due to the relatively small sample size, extensive data augmentation (including rotations and scaling) was used to expand the training set to 600 examples.
Results and Discussion
The Residual Attention U-Net significantly outperformed the baseline U-Net across all primary metrics. The model achieved a Dice Score (DSC) of 0.94, representing a 14.6% improvement over the standard U-Net's 0.82. Accuracy and precision also saw substantial gains, reaching 0.89 and 0.95, respectively.
Ablation studies confirmed that both the ResNeXt blocks and the LSH attention mechanism were vital to the model's success. The attention mechanism, in particular, acted as a critical filter, helping the decoder focus on relevant relative pixel positions to recover complex features that are typically lost during the down-sampling process in standard architectures.
Conclusion
This paper demonstrates that combining high-capacity feature extraction with memory-efficient attention mechanisms can provide a significant performance boost in medical imaging tasks. While the results are promising, the authors acknowledge that future work should involve validating the model on larger, more diverse cohorts to ensure generalization. Additionally, improving the interpretability of the model's internal decision-making process remains a priority for clinical adoption. The proposed framework offers a robust starting point for automated diagnostic tools that can adapt to new lung pathologies where ground-truth data may be scarce.
Summary updated