Abstract
In this paper, we present a hybrid deep learning framework named CTNet which combines convolutional neural network and transformer together for the detection of COVID-19 via 3D chest CT images. It consists of a CNN feature extractor module with SE attention to extract sufficient features from CT scans, together with a transformer model to model the discriminative features of the 3D CT scans. Compared to previous works, CTNet provides an effective and efficient method to perform COVID-19 diagnosis via 3D CT scans with data resampling strategy. Advanced results on a large and public benchmarks, COV19-CT-DB database was achieved by the proposed CTNet, over the state-of-the-art baseline approach proposed together with the dataset.
Introduction
While rRT-PCR remains the "gold standard" for COVID-19 diagnosis, its sensitivity can be limited by sample collection and transportation issues. Computed Tomography (CT) offers a precise 3D view of lung lesions, but the sheer volume of scans creates a massive bottleneck for radiologists. Previous deep learning attempts to automate this process faced hurdles such as dataset limitations leading to overfitting and preprocessing inefficiencies. Many models either required manual slice selection or attempted to process entire 3D volumes, which is computationally expensive and leads to unstable training due to small batch sizes.
Methodology
The proposed CTNet framework addresses these challenges by breaking the detection process into three distinct stages:
- Data Resampling Strategy: To standardize input and allow for larger batch sizes, CTNet resamples each 3D scan into a fixed number of slices (32). Uniform sampling is used for scans with more than 32 slices, while random oversampling is used for those with fewer.
- CNN Feature Extractor with SE Attention: Resampled slices pass through a multi-stage residual CNN integrated with Squeeze-and-Excitation (SE) attention modules. These modules help the network weigh the importance of different feature maps, focusing on discriminative infection indicators.
- Information Aggregation: Extracted features are processed by two parallel branches: a Transformer Branch using multi-head self-attention to model global 3D context across slices, and a Fully Connected (FC) Branch for standard feature classification. Their outputs are fused using Element-Wise Add (EWA) for the final prediction.
Experimental Setup
The study utilizes the COV19-CT-DB database, a large public benchmark containing over 5,000 chest CT scans. The model was trained on 855 scans and validated on 1,079 scans. These scans vary significantly in length (50 to 700 slices), providing a robust test for the resampling strategy.
Results and Discussion
CTNet achieved a macro F1 score of 88.23% on the validation set, representing an 18.23% improvement over the baseline approach (70.00%) provided with the dataset. Ablation studies confirmed the value of the hybrid approach: the fused model (CNN + Transformer + FC) outperformed the version using only the FC branch (87.96%).
Beyond accuracy, CTNet is highly efficient, processing a case in just 380ms. This speed, combined with the stability gained from the resampling strategy (allowing a batch size of 32), makes it a viable candidate for real-world clinical use.
Conclusion
CTNet demonstrates that combining the local feature extraction of CNNs with the global context modeling of Transformers provides a powerful tool for COVID-19 detection. While the choice of 32 slices is empirical and further cross-geographic validation is needed, the framework sets a new benchmark for automated 3D CT diagnosis. Future work may involve visualizing model attention to improve clinical interpretability.
Summary updated