Abstract
Since the outbreak of COVID-19, the rapid spread of the epidemic has challenged medical institutions worldwide. Automated segmentation of pathological regions in CT images can significantly assist clinicians in rapid diagnosis. This paper proposes a novel network structure, QC-HC U-Net, which combines residual and dense connections (Quick Connection) with Hypercolumns in the decoder. This architecture effectively avoids vanishing gradients while extracting richer features. Using a combination of lung tumor data for pre-validation and public COVID-19 datasets, the model achieved a Dice Similarity Coefficient (DSC) of 85.31% for COVID-19 segmentation, demonstrating high sensitivity and specificity without overfitting.
Introduction
While RT-PCR remains the diagnostic standard for COVID-19, its sensitivity can be limited, sometimes yielding false negatives in symptomatic patients. Lung CT scans offer a critical alternative, revealing characteristic "ground-glass opacities" (GGO). However, manual segmentation of these lesions is labor-intensive and prone to inter-observer variability. Standard deep learning models like the 3D U-Net often struggle with vanishing gradients in deep layers and spatial blurring during upsampling. The QC-HC U-Net is designed to overcome these hurdles by optimizing feature propagation and spatial reconstruction.
Architectural Innovation: QC-HC U-Net
The proposed QC-HC U-Net (Quick Connection-Hypercolumns U-Net) introduces two primary structural enhancements to the standard U-Net framework:
- Quick Connection (QC): This hybrid approach integrates the strengths of Residual connections (ResNet) and Dense connections (DenseNet). By allowing information to bypass layers efficiently, the network preserves feature integrity across depths without the prohibitive computational costs of traditional dense blocks. This ensures robust learning of complex pathological features.
- Hypercolumns (HC): Standard decoders typically rely on the final layer for prediction, which can lose fine-grained spatial detail. By adding Hypercolumns, the model aggregates feature maps from multiple decoder levels. This fusion of high-level semantic context and low-level spatial information results in significantly sharper and more accurate segmentation boundaries.
Methodology and Training
To address the challenge of limited medical data, the researchers employed a two-stage validation process and extensive data augmentation.
- Datasets: The model was first trained and tested on 63 cases from the MSD Lung Tumor dataset to verify its general feature extraction capabilities. Subsequently, it was applied to 20 public COVID-19 CT cases (comprising over 1,800 annotated slices).
- Data Augmentation: Techniques such as space expansion (rotation/mirroring), Gaussian noise, and patchwise training (160x160x80 blocks) were used to prevent overfitting and improve model generalization.
- Loss Function: A hybrid loss function combining Categorical Cross-Entropy and the Tversky index was utilized. The Tversky index is particularly effective for imbalanced data, allowing for fine-tuning the balance between false positives and false negatives in small lesion detection.
Results and Discussion
The QC-HC U-Net demonstrated a substantial performance leap over established benchmarks. In COVID-19 segmentation tasks, the model achieved a Dice Similarity Coefficient (DSC) of 85.31%, compared to the 67.3% seen in baseline models. Quantitative analysis showed a sensitivity of 83.60% and a specificity of 99.96%.
Qualitatively, the model produced outlines that closely matched expert annotations, effectively capturing both large infected regions and smaller, isolated opacities that previous models often missed or blurred. The integration of Hypercolumns proved vital in maintaining anatomical precision during the reconstruction phase.
Conclusion
The QC-HC U-Net offers a powerful tool for the automated analysis of COVID-19 CT scans. By optimizing the way features are connected and reconstructed, the model achieves high accuracy even with relatively small datasets. Future work will focus on expanding the training data and further optimizing the Quick Connection modules to reduce computational overhead, potentially enabling real-time clinical deployment on standard hospital hardware.
Summary updated