Evaluating Deep Neural Network Architectures with Transfer Learning for Pneumonitis Diagnosis

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Pulmonology

Background:

• Pneumonitis, an inflammation of lung air sacs, poses significant risks, especially to vulnerable populations.
• Accurate and timely diagnosis of pneumonitis from X-ray images presents a considerable challenge.
• Deep learning shows promise in medical image analysis, offering potential for improved diagnostic accuracy.

Purpose of the Study:

• To evaluate the efficacy of transfer learning in deep learning models for pneumonitis classification from chest X-rays.
• To compare the performance of various transfer learning models against established deep learning architectures.
• To identify the optimal model for accurate pneumonitis detection, facilitating early diagnosis.

Main Methods:

• Utilized the Kaggle chest X-ray dataset for model training and evaluation.
• Applied transfer learning by fine-tuning classification heads of models pre-trained on ImageNet.
• Employed basic data augmentation techniques to enhance model robustness.

Main Results:

• The DenseNet201 model demonstrated superior performance, achieving an AUROC score of 0.966 and a recall score of 0.99.
• Comparative analysis showed DenseNet201 outperforming other evaluated transfer learning models.
• Class activation maps were visualized to understand DenseNet201's predictive pattern recognition.

Conclusions:

• Transfer learning is an effective strategy to mitigate the resource demands of training deep learning models for pneumonitis detection.
• DenseNet201 shows significant potential as a reliable tool for classifying pneumonitis from chest X-rays.
• Further research and visualization of model interpretations can enhance clinical trust and application of AI in diagnosing lung inflammation.