Source-linked AI summary
A Light CNN for detecting COVID-19 from CT scans of the chest
Matteo Polsinelli, Luigi Cinque, Giuseppe Placidi
TL;DR
The paper addresses rapid COVID-19 CT screening under limited data and constrained computing resources by developing a lightweight SqueezeNet-based CNN. Its modified model discriminates COVID-19 from pneumonia and healthy CT images, achieving reported test performance with classification times suitable for medium-end computers.
Problem
COVID-19 screening needs faster, reliable imaging-based methods because available CT datasets are limited and conventional CNN inference can require substantial computational power.
Method
The study develops a light SqueezeNet-based CNN with modified Fire Modules, Bayesian hyperparameter optimization, and class activation mapping.
Results
83.00% accuracy, 85.00% sensitivity, 81.00% specificity, 81.73% precision, and 0.8333 F1Score were achieved on the test dataset.
Takeaways & Limitations
The proposed CNN is reported to be more efficient than more complex CNN designs, requiring 7.81 seconds per CT image on a medium-end laptop without GPU acceleration.
Takeaways & Limitations
Test-2 performance was limited by the difficulty of analysis without a larger image dataset, reaching 0.5024 accuracy and 0.1900 sensitivity for COVID-19 images.
Abstract
from arXiv · showhide
OVID-19 is a world-wide disease that has been declared as a pandemic by the World Health Organization. Computer Tomography (CT) imaging of the chest seems to be a valid diagnosis tool to detect COVID-19 promptly and to control the spread of the disease. Deep Learning has been extensively used in medical imaging and convolutional neural networks (CNNs) have been also used for classification of CT images. We propose a light CNN design based on the model of the SqueezeNet, for the efficient discrimination of COVID-19 CT images with other CT images (community-acquired pneumonia and/or healthy images). On the tested datasets, the proposed modified SqueezeNet CNN achieved 83.00\% of accuracy, 85.00\% of sensitivity, 81.00\% of specificity, 81.73\% of precision and 0.8333 of F1Score in a very efficient way (7.81 seconds medium-end laptot without GPU acceleration). Besides performance, the average classification time is very competitive with respect to more complex CNN designs, thus allowing its usability also on medium power computers. In the next future we aim at improving the performances of the method along two directions: 1) by increasing the training dataset (as soon as other CT images will be available); 2) by introducing an efficient pre-processing strategy.
1 introduction
The paper targets rapid, reliable COVID-19 CT screening despite limited datasets, overfitting risks, and the computational demands of conventional CNNs. It develops a lightweight SqueezeNet-based classifier for distinguishing COVID-19 from pneumonia and healthy CT images.
- Motivation: COVID-19 screening requires faster alternatives because RT-PCR can be unavailable in sufficient quantities and takes several hours.The paper motivates CT-based and automated screening to support rapid diagnosis and disease control.
- Motivation: CT imaging is presented as a valid COVID-19 detection alternative, with reported sensitivity up to 98% versus 71% for RT-PCR.Automated classification is motivated by the visual burden and misclassification risk involved in evaluating many CT images quickly.
- Challenges: CNN training is constrained by limited COVID-19 CT datasets, making overfitting reduction necessary to learn discriminative features rather than memorize scans.The paper also notes that pneumonia-related lung damage can make CT images difficult to classify correctly.
- Challenges: CNN inference commonly requires expensive GPU hardware, which may be unavailable in CPU-based clusters, hospitals, and developing countries.This motivates reducing the computational overhead of automatic systems.
- Approach: The study develops and tests a light SqueezeNet-based CNN to discriminate COVID-19 from community-acquired pneumonia and/or healthy CT images.Bayesian hyperparameter optimization and class activation mapping are used across two datasets to support training and inspect relevant image regions.
2 methods
The method organizes two CT datasets into separate and mixed arrangements, then evaluates lightweight SqueezeNet variants and a custom architecture. The proposed design modifies Fire Modules and combines early spatial information with later features while tuning training hyperparameters Bayesianly.
- Datasets organization: The study uses Zhao et al. and Italian CT datasets in separate and mixed arrangements for training, validation, and testing.The mixed arrangement is intended to increase the training dataset, while the separate arrangement keeps data sources distinct.
- CNN design: Batch normalization is inserted before Fire Module expansion to improve convergence and reduce overfitting while limiting computation overhead.The modified design also replaces ReLU activations with ELU activations.
- CNN design: The experiments compare SqueezeNet without bypasses, SqueezeNet with bypasses, and a further custom CNN based on the no-bypass design.The custom architecture adds a transposed convolution, skip connection, weighted summation, concatenation, and global average pooling.
- CNN design: The custom CNN weights early spatial information with later feature information before global average pooling to improve accuracy.Feature maps from the second Custom Fire Module are expanded, combined through a skip connection, and weighted using a 1×1 convolution.
3 results and discussion
The experiments identify the proposed CNN as the strongest configuration across the tested arrangements, while CNN-1 and CNN-2 show different sensitivity-specificity behavior. CAMs and efficiency comparisons further support CNN-2’s behavior and the light model’s practical use.
- Experiments organization and hyperparameter optimization: Experiment #4 achieved the highest observed and estimated accuracy in dataset arrangement 1 and remained the best experiment in arrangement 2.The experiments compared SqueezeNet variants, transfer learning, and the proposed CNN using Bayesian hyperparameter optimization.
- Experiments organization and hyperparameter optimization: The proposed CNN achieved better results than different configurations of the original SqueezeNet.
- Training, Validation and Test: CNN-1 reached higher specificity at 0.85, whereas CNN-2 reached higher sensitivity at 0.8500.CNN-1 was better at recognizing not COVID-19 images, while CNN-2 was better at recognizing COVID-19 images.
- Training, Validation and Test: CNN-1 achieved only 0.5024 accuracy on Test-2, with 0.80 precision and 0.1900 sensitivity for COVID-19 images.The authors state that analyzing Test-2 is difficult without a larger image dataset.
- Training, Validation and Test: CNN-2’s CAMs focused more appropriately on lung regions, while CNN-1 often activated outside the lungs or in small isolated zones.For COVID-19 images, CNN-2’s activations improved toward the diseased lung area; for non-COVID-19 images, it considered the whole lung region.
- Comparison with recent works: Compared with prior works, the proposed method performed better than especially in sensitivity, but method achieved better results.The authors caution that differing methods and training and test datasets make direct quantitative comparison difficult.
- Comparison with recent works: The proposed CNN classified one CT image in 1.25 seconds on a high-end workstation and 7.81 seconds on a medium-end laptop.The authors report that the method could classify about 4,300 images in approximately 9.32 hours on medium-end computers and had better sensitivity efficiency relative to parameter count.
4 conclusion
The proposed CNN discriminates COVID-19 from other CT images and outperforms original SqueezeNet on both dataset arrangements. Its future development targets improved performance through larger training datasets and efficient preprocessing while preserving efficiency.
- 83.00% accuracy, 85.00% sensitivity, 81.00% specificity, 81.73% precision, and 0.8333 F1Score were achieved on the test dataset.
- The proposed CNN outperformed the original SqueezeNet on both dataset arrangements.
- 7.81 seconds was required to classify a single CT image on a medium-end laptop, demonstrating efficiency with limited hardware resources.
- Future work will increase the training dataset as new CT images become available to improve accuracy, sensitivity, specificity, precision, and F1Score.
- Efficient preprocessing strategies will be studied to improve accuracy while reducing computational overhead and preserving efficiency.