Introduction
While Deep Learning (DL) has shown immense potential in medical imaging, its "black-box" nature remains a significant barrier to clinical adoption. In the context of COVID-19, a model might correctly identify a positive case, but without knowing which features (like ground-glass opacities or lesions) led to that conclusion, radiologists cannot fully trust the output. Furthermore, training models on 3D CT volumes is challenging because labels are often provided at the patient level, not the individual image slice level. Since many slices in a COVID-positive patient's scan may appear normal, simply applying a patient-level label to every slice introduces significant "noise" into the training process.
Methods
The researchers proposed a three-pronged architectural approach to solve these issues, focusing on making the model's "attention" a core part of its architecture rather than a post-processing afterthought.
The Explainable Module
Standard Class Activation Mapping (CAM) is usually a post-processing step. The authors replaced the traditional Global Average Pooling (GAP) and Fully Connected (FC) layers with a 1×1 convolution layer. This mathematical equivalence allows the model to learn activation maps adaptively during training.
Slice Integration Module
To handle the noise of patient-level labels, the team developed a strategy that divides a 3D CT volume into disjoint sections. It calculates the probability of infection for each section based on the most "suspicious" slices and then aggregates these into a final patient-level score. This prevents healthy slices from "diluting" the training signal of infected ones.
Local and Global Interpretability
The framework incorporates LIME (Local Interpretable Model-agnostic Explanations) to highlight specific "super-pixels" that contribute to a diagnosis and SHAP (SHapley Additive exPlanation) values to ensure the model's interpretations remain consistent and theoretically sound.
Datasets
- Private Dataset: Used for initial training, consisting of 804 CT volumes (380 COVID-positive, 424 COVID-negative) collected from four hospitals in China.
- CC-CCII Dataset: A large, publicly available dataset containing 2,034 CT volumes used to benchmark the model against other state-of-the-art architectures.
Key Results
The proposed model significantly outperformed existing state-of-the-art (SOTA) COVID-19 classifiers on the CC-CCII dataset, achieving a state-of-the-art accuracy of 89.23%.
| Method | Accuracy (%) | Precision (%) | Recall (%) | AUC (%) |
|---|---|---|---|---|
| ResNet-50 | 53.72 | 61.42 | 77.20 | 46.30 |
| COVID-Net | 53.65 | 61.35 | 77.20 | 44.53 |
| VB-Net | 76.73 | 85.26 | 77.58 | 89.48 |
| Ours | 89.23 | 89.98 | 93.86 | 93.22 |
Qualitatively, the model's activation maps consistently highlighted the same lesion areas that a radiologist would look for, such as peripheral lung opacities.
Discussion and Future Work
By moving the explanation mechanism into the forward pass of the network, the model is forced to learn features that are inherently more interpretable. Using LIME to group pixels into "super-pixels" further allows the model to provide explanations that align with human perception—identifying regions and textures rather than individual, disconnected pixels.
However, the authors note that this study is an "initial comparison." Further testing is needed to see if this specific XAI architecture holds up across other lung diseases or different imaging modalities like MRI. Future work will also involve "human-in-the-loop" testing to determine if these visual explanations actually improve the speed or accuracy of real-world radiologist diagnoses.
Summary updated