Lung Nodule Sizes Are Encoded When Scaling CT Image for CNN's

Area of Science:

• Medical Imaging
• Radiomics
• Artificial Intelligence

Background:

• Noninvasive diagnosis of lung cancer in early stages is crucial.
• Radiomics offers potential for early lung cancer detection.
• Nodule size is a key indicator of malignancy in clinical practice.

Purpose of the Study:

• To evaluate the ability of convolutional neural networks (CNNs) to capture nodule size in computed tomography (CT) images after image resizing.
• To assess the generalizability of CNNs in size classification across different image datasets.

Main Methods:

• Utilized the National Lung Screening Trial dataset for lung nodule analysis.
• Extracted lung nodule patches and resized them to 100x100 pixels for CNN input.
• Validated findings using the Common Objects in Context (COCO) dataset with categories like bears, cats, and dogs.

Main Results:

• CNNs achieved high accuracy in classifying lung nodules into small and large size groups after resizing.
• Cross-validation experiments on the COCO dataset demonstrated high performance (AUCs ranging from 0.952 to 0.979) for size classification.
• Rescaling images for CNN input enables accurate object size detection.

Conclusions:

• CNNs can effectively determine lung nodule size from rescaled CT images, supporting early lung cancer diagnosis.
• The ability of CNNs to discern object size is not limited to medical images and applies to general object recognition.
• Publicly available source code facilitates further research and development in this area.