Abstract

This study investigates the use of deep learning models—ResNet, DenseNet, and VGG—for detecting COVID-19 in chest CT scans. The goal was to evaluate how accurately these models can classify CT scan images as either positive or negative for COVID-19.

The models were trained and tested on a dataset of labeled chest CT scans, and their performance was measured using accuracy, precision, recall, F1-score, and AUC-ROC. Results showed that DenseNet performed the best, with an accuracy of 94.2%, precision of 93.5%, recall of 95.3%, F1-score of 94.4%, and an AUC-ROC of 0.97.

This study highlights the potential of using deep learning models, especially DenseNet, as a tool to assist healthcare professionals in diagnosing COVID-19. However, further clinical testing is needed before these models can be used in real-world healthcare settings.

Keywords: COVID-19, deep learning, ResNet, DenseNet, VGG, CT scans, diagnostic tool, AI in healthcare.

Problem Statement

COVID-19 continues to pose significant challenges to healthcare systems worldwide, requiring timely and accurate detection methods to manage the pandemic. Current diagnostic methods such as PCR tests are resource-intensive, slow, and sometimes unavailable in certain regions. CT scans, which provide detailed imaging of the lungs, are being explored as a faster diagnostic tool. However, the manual interpretation of these scans by radiologists is a complex and time-consuming task. Therefore, there is a need to develop automated systems for detecting COVID-19 from CT scans using deep learning models, which could reduce the workload on healthcare professionals and improve diagnostic accuracy and speed.

Existing System

1. Manual CT scan analysis by radiologists
2. RT-PCR test for confirmation
3. Time-consuming and dependent on expert availability
4. Higher chance of human error
5. Limited automation

Proposed System

1. Use pre-trained deep learning models (ResNet, DenseNet, VGG)
2. Automatic feature extraction from CT images
3. Faster prediction with high accuracy
4. Real-time classification support
5. Assist doctors in early detection

Software Requirements

1. Python
2. TensorFlow / PyTorch
3. OpenCV
4. NumPy, Pandas
5. Jupyter Notebook / VS Code
6. Matplotlib / Seaborn

Hardware Requirements

1. Minimum 8GB RAM (16GB recommended)
2. GPU (NVIDIA CUDA supported – recommended for training)
3. Intel i5/i7 or equivalent processor
4. 256GB storage

Advantages

1. Faster diagnosis
2. High accuracy using deep learning
3. Reduces workload of radiologists
4. Automated feature extraction
5. Scalable for large datasets

Disadvantages

1. Requires large labeled dataset
2. High computational cost
3. GPU dependency for training
4. Risk of overfitting
5. Not a complete replacement for medical experts