Automated Multi-class Classification of Lung Diseases from CT-scans Using Computer Vision

Abstract

Background: Computed tomography (CT) scans and X-rays play a crucial role in diagnosing lung diseases. However, challenges such as imbalanced dataset and low image contrast can deter the performance of diagnostic models. To address these issues, convolutional neural networks (CNNs) offer a promising solution for respiratory diseases. Previous research in this domain has been conducted on smaller dataset, typically limited to just three to four classes. This study aimed to create and assess a CNN-based framework for the classification of multi-class lung diseases using a largescale CT dataset, and to compare its performance with established pertained CNN models. Methods: A data set of 466,476 CT images was constructed, representing nine distinct lung disease groups. Data augmentation and histogram equalization were used to improve diversity and contrast. The proposed CNN model, along with pretrained networks (AlexNet, SqueezeNet, and GoogleNet), was trained by dividing the whole dataset into 80% for training and the remaining 20% for testing. Performance and results were evaluated using parameters such as accuracy, sensitivity, and specificity. Results: The proposed CNN consistently outperformed pretrained models across nine experiments. It achieved a perfect classification performance of 100% accuracy, sensitivity, and specificity in one experimental setting, while AlexNet achieved the best results among pretrained models with 99.21% accuracy. Overall, the proposed method demonstrated superior robustness and reliability across all disease categories. Conclusion: The proposed CNN-based approach provides a highly effective method for multi-class lung disease classification from CT scans. Its exceptional performance highlights its potential for integration into clinical practice, offering improved diagnostic accuracy and efficiency.