Source-linked AI summary
Explainable COVID-19 Detection Using Chest CT Scans and Deep Learning
Hammam Alshazly, Christoph Linse, Erhardt Barth, Thomas Martinetz
TL;DR
The paper addresses automated COVID-19 detection from chest CT images, motivated by limitations of conventional testing and manual image review. It evaluates advanced CNNs with architecture-specific transfer learning on two CT datasets and adds visual explanations. The best models achieve high performance, including 99.4% accuracy and 99.8% sensitivity on the SARS-CoV-2 dataset, while visualization results show separated feature clusters and localized COVID-19 regions.
Problem
Manual CT interpretation is time-consuming, while RT-PCR testing can face shortages, long turnaround times, and false negatives.
Method
The study compares advanced CNN architectures using transfer learning with custom-sized inputs selected for each architecture, then applies t-SNE and Grad-CAM for visual explanations.
Results
99.4% average accuracy and 99.8% average sensitivity are achieved on the SARS-CoV-2 CT dataset by the best reported models.
Takeaways & Limitations
The visualizations show separated COVID-19 and Non-COVID-19 feature clusters and localize COVID-19-associated regions in CT images.
Takeaways & Limitations
Performance is constrained on COVID19-CT by cross-source heterogeneity, overlapping lung-disease findings, and variations in contrast and spatial resolution.
Abstract
from arXiv · showhide
This paper explores how well deep learning models trained on chest CT images can diagnose COVID-19 infected people in a fast and automated process. To this end, we adopt advanced deep network architectures and propose a transfer learning strategy using custom-sized input tailored for each deep architecture to achieve the best performance. We conduct extensive sets of experiments on two CT image datasets, namely the SARS-CoV-2 CT-scan and the COVID19-CT. The obtained results show superior performances for our models compared with previous studies, where our best models achieve average accuracy, precision, sensitivity, specificity and F1 score of 99.4%, 99.6%, 99.8%, 99.6% and 99.4% on the SARS-CoV-2 dataset; and 92.9%, 91.3%, 93.7%, 92.2% and 92.5% on the COVID19-CT dataset, respectively. Furthermore, we apply two visualization techniques to provide visual explanations for the models' predictions. The visualizations show well-separated clusters for CT images of COVID-19 from other lung diseases, and accurate localizations of the COVID-19 associated regions.
1 Introduction
The paper motivates automated COVID-19 detection from chest CT images and evaluates advanced deep CNNs with transfer learning and visualization-based explanations. Its strategy uses architecture-specific inputs and visual analyses to improve detection performance and transparency.
- RT-PCR shortages, high turnaround times, and false negatives motivate additional tools for identifying COVID-19 patients.
- Manual radiologist examination of chest CT scans can be time-consuming and impractical in emergencies, motivating automated detection tools.
- t-SNE reveals separated COVID-19 and Non-COVID-19 feature clusters, while Grad-CAM localizes COVID-19-associated regions annotated by expert radiologists.
- The study compares 12 deep CNNs with different architectural designs and depths for COVID-19 detection from chest CT images.
- The proposed transfer-learning strategy fine-tunes networks with custom-sized inputs selected for each architecture and uses the LAMB optimizer.
- 99.4% average accuracy and 99.8% average sensitivity are achieved by the best models across the two CT datasets.
2 Related Work
Prior work established deep learning as a prominent approach for automated COVID-19 detection from medical images. This paper builds on that literature by comparing advanced CNNs on two large CT datasets and adding visual explanations.
- Existing COVID-19 imaging studies use either newly designed deep networks or established CNN architectures for automated detection.
- Earlier research demonstrated the potential of chest CT and deep learning for distinguishing COVID-19 pneumonia from other pneumonia types.
- Handcrafted features and conventional classifiers have also been applied, with one reported classification scheme achieving an F1-score of 83%.
- This paper compares 12 deep networks on two large CT image datasets using standard evaluation metrics and provides visualizations for explainability.
3 Deep Network Architectures
The paper evaluates diverse deep CNN architectures for COVID-19 detection from chest CT scans, spanning efficiency-focused, multi-branch, residual, separable-convolution, and channel-shuffling designs.
- The study considers SqueezeNet, Inception, ResNet, ResNeXt, Xception, ShuffleNet, and DenseNet as advanced CNN architectures for CT-based COVID-19 identification.
- SqueezeNet: SqueezeNet uses fire modules that reduce dimensionality with 1 × 1 squeeze filters before expanding through 1 × 1 and 3 × 3 convolutions.
- Inception: Inception V3 uses branching modules with different convolutional kernel sizes and pooling to capture patterns at multiple scales.
- ResNet: ResNet uses skip connections in residual modules, with basic blocks for ResNet18 and bottleneck blocks for ResNet50 and ResNet101.
- ResNeXt: ResNeXt applies identical transformations across branches, with cardinality controlling the number of branches whose outputs are aggregated.
- Xception and ShuffleNet: Xception separates spatial and cross-channel correlations using depthwise separable convolutions, while ShuffleNet reduces computation through grouped convolutions and channel shuffling.
4 Transfer Learning
The paper uses transfer learning to adapt pretrained networks to scarce CT-image data, fine-tuning all weights while resizing images for the required network inputs.
- Transfer learning initializes models with networks trained on millions of images for tasks with only hundreds or thousands of examples.
- The study uses fine-tuning, optimizing all network weights for the new COVID-19 detection task rather than training only newly added layers.
- CT images are replicated across the R, G, and B channels because the pretrained networks require RGB input.
- Images with varying spatial sizes are scaled to meet the input requirements of the selected architectures.
5 Experiments and Results
The experiments evaluate 12 deep CNNs on two CT datasets, combining quantitative detection results with t-SNE and Grad-CAM explanations. Performance is strongest on SARS-CoV-2 CT, while COVID19-CT shows lower and more variable results because of heterogeneity and limited sample size.
- Experimental design: The study evaluates fine-tuned deep networks on two CT datasets and reports quantitative results, confusion matrices, and visual explanations.The experiments also localize COVID-19-associated regions using Grad-CAM.
- SARS-CoV-2 CT results: 99.4% average accuracy and 99.4% F1-score make ResNet101 the strongest overall model on the SARS-CoV-2 CT dataset.Its average sensitivity is 99.1% and specificity is 99.6%.
- SARS-CoV-2 CT results: 99.8% average sensitivity is achieved by InceptionV3 on the SARS-CoV-2 CT dataset, with one COVID-19 image falsely predicted as negative on average.SqueezeNet has the lowest performance, while ShuffleNet improves approximately 2% over it across metrics and is more computationally efficient.
- Dataset effects: COVID19-CT performance is lower because cross-source heterogeneity, overlapping disease manifestations, contrast variation, and spatial-resolution differences hinder discriminative feature extraction.The dataset’s smaller folds also produce stronger inter-fold variation, particularly for specificity.
- COVID19-CT results: 92.9% average accuracy and 93.7% average sensitivity make DenseNet201 the best model on the COVID19-CT dataset.DenseNet169 achieves 91.6% average accuracy and the same sensitivity as the best model.
- Visual explanations: t-SNE produces well-separated COVID-19 and Non-COVID-19 feature clusters, while Grad-CAM localizes regions marked by expert radiologists.On external and partial-lung CT images, InceptionV3 correctly classifies COVID-19 cases and localizes disease-related abnormalities.
6 Conclusion
The paper reports strong COVID-19 detection from chest CT images using deep learning, with high performance across two datasets and visualizations that explain learned features and localized abnormalities.
- Custom-sized inputs and extensive experiments with advanced deep network architectures were used to improve COVID-19 detection performance.The study considered variants of advanced architectures and determined input sizes for each network.
- t-SNE visualizations showed well-separated COVID-19 and Non-COVID-19 clusters in both the SARS-CoV-2 and COVID19-CT datasets.The SARS-CoV-2 visualization distinguishes train and test samples by color within the two classes.
- Grad-CAM visualizations localized COVID-19-related regions in correctly classified images from both datasets.InceptionV3 was shown for the SARS-CoV-2 dataset, while DenseNet169 was shown for the COVID19-CT dataset.
- External CT images from different publications were correctly classified as COVID-19, with associated abnormalities accurately localized according to expert markings.The tested images included different COVID-19 manifestations marked by specialists or radiologists.