Deep CNN models for pulmonary nodule classification: Model modification, model integration, and transfer learning

Area of Science:

• Medical Imaging Analysis
• Artificial Intelligence in Healthcare

Background:

• Deep learning excels in natural image analysis but faces challenges in medical imaging due to data limitations.
• Accurate classification of pulmonary nodules is crucial for diagnosing lung cancer.

Purpose of the Study:

• To explore strategies for utilizing deep convolutional neural networks (CNNs) for classifying malignant and benign pulmonary nodules in CT images.
• To compare different approaches for optimizing CNN performance in medical image analysis.

Main Methods:

• Experiments utilized the LIDC-IDRI public database (1018 cases).
• Three strategies were implemented: CNN architecture modification, CNN integration, and transfer learning.
• Eleven deep CNN models were evaluated on the same dataset.

Main Results:

• A modified CifarNet achieved an AUC of 0.90.
• Integrated CNN models offered reduced complexity without significant performance gains.
• Transfer learning, especially fine-tuned ResNet, yielded the best results with an AUC of 0.94, 91% sensitivity, and 88% overall accuracy.

Conclusions:

• Model modification, integration, and transfer learning are effective for optimizing deep CNNs in pulmonary nodule classification.
• Transfer learning is the preferred strategy for applying deep learning to medical imaging tasks.