Convolutional neural network-based program to predict lymph node metastasis of non-small cell lung cancer using...

Area of Science:

• Medical Imaging
• Artificial Intelligence in Oncology
• Radiology

Background:

• Non-small cell lung cancer (NSCLC) diagnosis relies on accurate staging, particularly lymph node metastasis assessment.
• Positron emission tomography (PET) scans using 2-deoxy-2-[F-18]fluoro-D-glucose (FDG) are crucial for staging NSCLC.
• Maximum intensity projection (MIP) images derived from FDG-PET scans offer a summarized view for analysis.

Purpose of the Study:

• To develop and evaluate a convolutional neural network (CNN) for predicting lymph node metastasis in NSCLC using FDG-PET MIP images.
• To assess the CNN's effectiveness as a diagnostic aid for radiologists in this task.
• To quantify improvements in diagnostic accuracy and efficiency.

Main Methods:

• A ResNet-50 based CNN was trained on FDG-PET MIP images from 304 NSCLC patients.
• The CNN's performance was evaluated on a separate test set of 131 patients.
• Seven radiologists independently interpreted the test set images twice: once unaided and once with CNN assistance.

Main Results:

• The CNN achieved 0.748 predictive accuracy for lymph node metastasis.
• CNN assistance significantly reduced prediction error for 6 out of 7 radiologists.
• Radiologist reading time was significantly reduced by a median of 38.0% for 5 out of 7 radiologists.

Conclusions:

• A CNN-based program shows potential to assist radiologists in predicting NSCLC lymph node metastasis.
• The tool can enhance diagnostic confidence and decrease reading time without compromising accuracy.
• Further validation is needed, especially with MIP images in limited settings.