LDDNet: A Deep Learning Framework for the Diagnosis of Infectious Lung Diseases

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Science

Background:

• Lung diseases, including COVID-19 and pneumonia, pose significant global health challenges.
• Accurate and timely diagnosis from medical imaging (CT scans, X-rays) is crucial for effective treatment.
• Existing deep learning models require optimization for improved lung disease classification performance.

Purpose of the Study:

• To propose and evaluate a novel deep learning framework, LDDNet, for analyzing lung diseases from chest CT scans and X-ray images.
• To enhance the DenseNet201 architecture for superior performance in detecting COVID-19 and pneumonia.
• To compare LDDNet's diagnostic accuracy against established deep learning models.

Main Methods:

• Developed LDDNet by augmenting the DenseNet201 model with additional pooling, dense, dropout layers, and batch normalization.
• Trained and validated LDDNet on three distinct datasets: one CT scan dataset (1043 images) and two X-ray datasets (5935 and 5002 images).
• Optimized model hyperparameters and evaluated performance using Adam, Nadam, and SGD optimizers, with Nadam yielding the best results.

Main Results:

• LDDNet achieved high accuracy in classifying COVID-19 from CT scans (99.36%) and X-rays (99.55% on an imbalanced dataset, 97.07% on a balanced dataset).
• The Nadam optimizer consistently provided superior results across both CT and X-ray datasets.
• LDDNet demonstrated performance exceeding that of ResNet152V2 and XceptionNet on the tested lung disease datasets.

Conclusions:

• The proposed LDDNet framework offers a highly accurate and efficient deep learning solution for diagnosing lung diseases from medical imaging.
• LDDNet's optimized architecture and hyperparameter tuning contribute to its superior performance compared to existing methods.
• This framework holds potential for improving early detection and management of respiratory illnesses like COVID-19 and pneumonia.