Abstract
COVID-19 has significantly impacted global healthcare systems. While RT-PCR is the gold standard for diagnosis, its limited sensitivity and labor-intensive nature hinder mass screening. This paper proposes an automated detection system using chest CT scans. The framework employs an ensemble of three transfer learning-based CNNs—VGG-11, Wide ResNet-50-2, and Inception v3. A novel fuzzy rank-based fusion strategy using the Gompertz function adaptively combines decision scores. Evaluated on two public datasets, the model achieves state-of-the-art performance, offering a reliable tool for rapid COVID-19 screening.
Introduction
The ongoing challenges of the COVID-19 pandemic, exacerbated by emerging variants, necessitate early and accurate detection. RT-PCR tests, while standard, exhibit a sensitivity of only approximately 71% and require significant time and labor. Chest CT scans provide a faster alternative with distinct pathological patterns, such as ground-glass opacities. However, manual interpretation by radiologists is time-consuming. While individual AI models have been explored, they often lack the robustness needed for clinical variation. This research addresses these gaps by proposing a multi-model ensemble approach to enhance diagnostic accuracy and reliability.
Methodology
The proposed framework operates in two distinct stages: feature extraction and decision fusion.
Feature Extraction via Transfer Learning
The system utilizes three diverse Convolutional Neural Network (CNN) architectures: VGG-11, Wide ResNet-50-2, and Inception v3. These models, pre-trained on ImageNet, are fine-tuned on COVID-19 specific CT data. By using multiple architectures, the system captures a broader spectrum of spatial features than any single model could achieve alone.
Fuzzy Rank-Based Fusion
The core innovation lies in the fusion mechanism. Rather than simple averaging, the authors implement a re-parameterized Gompertz function to assign "fuzzy ranks" to model predictions. The Gompertz function, typically used to model mortality or growth saturation, maps confidence scores to ranks non-linearly. This allows the system to calculate a Final Decision Score (FDS) that adaptively weights the most "certain" models for each specific image, effectively handling the uncertainty inherent in medical imaging.
Experimental Setup
The model was validated using two major public datasets:
- SARS-COV-2 CT-scan dataset: 1,252 COVID-positive and 1,229 non-COVID scans.
- Harvard Dataverse chest CT dataset: 2,167 COVID-positive and 2,005 non-COVID scans. Both datasets were partitioned into 70% training and 30% testing sets to evaluate generalization performance.
Key Results and Insights
The ensemble method consistently outperformed individual base models and standard ensemble techniques (e.g., majority voting).
- High Accuracy: The system achieved 98.93% accuracy on the SARS-COV-2 dataset and 98.80% on the Harvard Dataverse dataset.
- Superior Sensitivity: With a sensitivity of 98.93%, the model significantly exceeds the ~71% sensitivity of RT-PCR, making it a powerful tool for reducing false negatives.
- Statistical Significance: McNemar’s test confirmed that the performance improvements were statistically significant (p < 0.05) compared to individual base models.
- Low False Positives: ROC curves with an AUC of ~0.98 demonstrate a very low false-positive rate, essential for minimizing unnecessary clinical burden.
The Gompertz Advantage
The choice of the Gompertz function allows for a sophisticated "refereeing" process. By modeling classification error non-linearly, the system provides "adaptive priority," evaluating classifier confidence on a per-sample basis. Interestingly, the researchers found that the simpler VGG-11 outperformed the deeper VGG-19, suggesting that for specialized medical datasets, excessive model complexity can lead to overfitting.
Conclusion and Future Work
While highly accurate, the ensemble approach introduces computational overhead. Future research may explore "snapshot ensembling" to reduce processing requirements. Additionally, incorporating lung segmentation could further improve focus on relevant pathological features. The authors envision this framework as a "plug-and-play" clinical tool, providing rapid, reliable second opinions to healthcare professionals in high-pressure environments.
Source code is available on GitHub.
Summary updated