Abstract
For evaluation, we constructed a dataset of 258 chest CT volumes with 153 VP, 42 BP, and 63 normal control cases, for a total of 43,421 slices. We implemented a comprehensive comparison between our SLP-Net and several state-of-the-art methods across the dataset. Our proposed method obtained significant performance without a large amount of data, outperformed other slice-based and volume-based approaches. The superior evaluation performance achieved in the classification experiments demonstrated the ability of our model in the differential diagnosis of VP, BP and normal cases.
Introduction
In the wake of the COVID-19 pandemic, AI development for lung disease detection has expanded rapidly. However, most systems focus narrowly on identifying viral pneumonia (VP) while often ignoring bacterial pneumonia (BP). Clinically, this is a major hurdle: VP and BP often present with similar "ground-glass opacities" on a CT scan, yet they require vastly different treatments—antibiotics for BP versus supportive care for VP.
Current AI approaches generally fall into two categories: 2D slice-based methods, which lose spatial continuity, and 3D volume-based methods, which are computationally heavy and require massive training data. The authors propose SLP-Net as a "Goldilocks" solution: a model that captures spatial context without the overhead of full 3D convolutions.
Methodology: SLP-Net Architecture
The core of this research is SLP-Net (Sequence-based Lung Pneumonia classification Network). It treats chest CT scans as a time sequence rather than a static 3D volume or independent 2D slices.
The architecture is built on three pillars:
- Res2Net Encoder: Extracts features at multiple scales within a single layer, crucial for detecting lesions of varying sizes.
- ConvLSTM (Convolutional Long Short-Term Memory): Acts as the "memory" of the network. By treating a sequence of CT slices like frames in a video, it captures how features evolve along the axial dimension while preserving spatial information.
- Adaptive-Weighted Cross-Entropy (ACE) Loss: A novel loss function that ensures the model effectively accumulates and encodes valid features into the final prediction without "forgetting" earlier slices.
The system also utilizes Grad-CAM attention maps to ensure that identified pneumonia "hotspots" in adjacent slices overlap, thereby reducing false positives.
Experimental Results
The researchers evaluated SLP-Net on a dataset of 258 chest CT volumes (43,421 slices), including 153 VP cases, 42 BP cases, and 63 normal controls.
SLP-Net consistently outperformed state-of-the-art 2D (e.g., ResNet34, Xception) and 3D models (e.g., C3D, I3D):
- Accuracy: Achieved a Balanced Accuracy (B-Acc) of 95.23% and a Kappa score of 0.9263.
- Efficiency: Maintained the smallest model size (34.4 MB) and faster inference times (~39.5 ms) compared to 3D architectures.
- Optimal Sequence: A sequence of 10 slices with a 5-slice overlap was found to be the optimal configuration for balancing context and training sample volume.
Discussion and Conclusion
By treating the Z-axis of a CT scan as a "time" dimension, the researchers achieved the context of 3D imaging with the efficiency of 2D imaging. The ConvLSTM is specifically tuned to pick up on the frequency and distribution patterns that differentiate VP from BP.
While the model shows great promise, future work includes external validation on multi-center data to ensure robustness across different imaging protocols. The authors suggest this sequence-based approach could be adapted for other conditions where spatial patterns across slices are key, such as lung cancer or neurodegenerative diseases.
Summary updated