LUNGx Challenge for computerized lung nodule classification.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Radiology

Background:

• Accurate classification of lung nodules on computed tomography (CT) scans is crucial for patient management.
• Computerized methods offer potential for improving the efficiency and accuracy of lung nodule classification.

Purpose of the Study:

• To describe the LUNGx Challenge, a competition for developing computerized methods to classify lung nodules as benign or malignant.
• To compare the performance of these computerized methods against human radiologists on the same dataset.

Main Methods:

• The LUNGx Challenge provided standardized calibration and testing datasets of CT scans with 73 lung nodules (37 benign, 36 malignant).
• Ten research groups submitted computerized methods for nodule classification.
• Six radiologists participated in an observer study performing the same classification task.

Main Results:

• Computerized methods achieved Area Under the Receiver Operating Characteristic Curve (AUC) values ranging from 0.50 to 0.68, with only three outperforming random guessing.
• Radiologists achieved AUC values from 0.70 to 0.85.
• Three radiologists performed statistically better than the best-performing computer method.

Conclusions:

• Current computerized methods for lung nodule classification on CT scans lag behind radiologist performance.
• The LUNGx Challenge dataset will serve as a valuable resource for advancing AI in medical imaging research.
• Further development is needed to improve the accuracy of AI in differentiating benign from malignant lung nodules.