Discriminating TB lung nodules from early lung cancers using deep learning.

Area of Science:

• Radiology
• Artificial Intelligence
• Oncology

Background:

• Developing countries face challenges in differentiating lung cancer from latent tuberculosis (LTB) due to high smoking rates and endemic TB.
• Nodules from LTB can mimic lung cancer on CT scans, complicating early diagnosis.
• Distinguishing between lung cancer and LTB often requires invasive procedures with associated risks.

Purpose of the Study:

• To develop and evaluate a deep convolutional neural network (DNN) for noninvasive differentiation of lung cancer from LTB nodules.
• To assess the DNN's performance using CT images from established lung screening and TB research datasets.

Main Methods:

• A customized 15-layer, 2D deep convolutional neural network (DNN) architecture based on VGG16 was employed.
• Transfer learning was utilized, with the DNN trained and tested on CT images from the National Lung Screening Trial and TB Portals.
• Performance was evaluated under locked and step-wise unlocked pretrained weight conditions.

Main Results:

• The DNN, utilizing unlocked pretrained weights, achieved a high accuracy of 90.4%.
• An F score of 90.1% was recorded, indicating robust performance in classification.
• The model demonstrated significant capability in distinguishing between lung cancer and LTB.

Conclusions:

• The developed DNN shows potential as a noninvasive screening tool for lung cancer and LTB.
• This AI-driven approach can reliably detect and differentiate between these conditions, reducing the need for invasive procedures.
• The findings support the clinical utility of DNNs in complex diagnostic scenarios involving pulmonary nodules.