NLSTseg: A Pixel-level Lung Cancer Dataset Based on NLST LDCT Images.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Low-dose computed tomography (LDCT) is crucial for early lung cancer detection.
• Computer-Aided Diagnosis (CAD) systems assist radiologists but require extensive annotated datasets.
• Existing public datasets with pixel-level segmentation annotations for LDCT images are limited.

Purpose of the Study:

• To develop a novel, publicly available dataset with pixel-level annotations of lung lesions from LDCT images.
• To address the scarcity of annotated data for training AI-driven CAD systems in lung cancer diagnosis.

Main Methods:

• Utilized National Lung Screening Trial (NLST) LDCT images.
• Annotated 715 lung lesions (662 tumors, 53 nodules) across 605 patient scans.
• Included detailed lesion volume and location information.

Main Results:

• The dataset comprises LDCT scans from 605 patients with 715 annotated lung lesions.
• Lesion volumes varied significantly, with a majority (500) smaller than 5 cm³.
• A 2D U-Net model trained on this dataset achieved a 0.95 Intersection over Union (IoU) on the training set.

Conclusions:

• The developed dataset significantly enhances the diversity and usability of lung cancer annotation resources.
• This resource is expected to accelerate the development and clinical adoption of AI-based CAD systems for lung cancer screening.
• Facilitates improved accuracy and efficiency in interpreting LDCT scans for early lung cancer detection.