Classification of Benign and Malignant Lung Nodules Based on Deep Convolutional Network Feature Extraction

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Computed tomography (CT) imaging is crucial for early lung nodule diagnosis.
• Accurate classification of lung nodule nature (benign vs. malignant) is challenging due to radiologist subjectivity.
• Deep convolutional neural networks (CNNs) show promise for lung nodule classification with increasing image data.

Purpose of the Study:

• To develop a novel deep convolutional network approach for accurate benign and malignant lung nodule classification.
• To address the challenge of gradient dispersion in deep networks during training.
• To provide an objective and efficient tool for aiding radiologists in lung nodule diagnosis.

Main Methods:

• Lung nodule images were segmented, extracted, and preprocessed using zero-phase component analysis whitening.
• A deep convolutional network was constructed by integrating a multilayer perceptron.
• The network was fine-tuned using minibatch stochastic gradient descent with a momentum coefficient to mitigate gradient dispersion.

Main Results:

• The proposed deep convolutional network achieved a classification accuracy of 96.0% on a dataset of 750 lung nodules.
• The fine-tuning method effectively avoided gradient dispersion during network training.
• Experimental verification confirmed the method's efficacy in classifying lung nodules.

Conclusions:

• The developed deep convolutional network offers a highly accurate and objective method for classifying lung nodules.
• This approach can serve as an efficient decision support tool for radiologists in medical image analysis.
• The study highlights the potential of advanced deep learning techniques in improving lung cancer diagnosis.