In our study, we evaluated the performance of a commercially available AI algorithm (Lunit INSIGHT for CXR, accessible at https://insight.lunit.io) for nodule detection using the publicly available National Lung Screening Trial (NLST) data set. In this repository, we publicly release the nodule label used in our study. To download the original DICOM file corresponding to our label, request should be made to CDAS (https://cdas.cancer.gov/).
The NLST was a clinical trial conducted to compare LDCT and standard chest X-ray as screening tools to reduce mortality from lung cancer. The data set consists of anonymized DICOM files of screening chest X-rays, cancer diagnosis and survival information, and NLST radiologists' assessment of the chest X-rays. A more detailed overview and study design of NLST can be found in the Radiology paper (https://pubs.rsna.org/doi/full/10.1148/radiol.10091808).
A subset (n=577) of the patients with available baseline images were selected and annotated for the presence of nodules. The selection process was as belows:
- CXRs of patients (n=48) with lung cancer diagnosis within 1 year of T0 screening were selected.
- CXRs of patients (n=50) with all three screening CXRs and non-calcified nodules (as marked by NLST radiologists), but without a lung cancer diagnosis, were selected.
- CXRs of patients (n=480) not satisfying the above two criteria were selected
- CXR that did not contain adequate view of the lung (n=1) was removed from analysis.
The nodule data set were labeled by two radiologists (K.H.K., with 6 years of experience, and M. K., with 21 years of experience).
- Each radiologist independently evaluated T0 CXRs for the presence of non-calcified nodules that are ≥4mm.
- Cancer characteristics as well as all available sequential CXRs for the patient and NLST radiologist labels for the CXRs were provided during annotation.
- A CXR was labeled as positive for non-calcified nodules if the two radiologists agreed on the presence of non-calcified nodules and negative, if otherwise.
- A CXR was labeled as positive for malignant pulmonary nodules if it was positive for both lung cancer and non-calcified nodules.
cancer_pathology
: whether patient was diagnosed with lung cancer within 1 year of baseline imaging.cancer_time
: time interval from baseline imaging to cancer diagnosis.lesion_size_at_Dx
: size (mm) of the lesion at diagnosis.location_at_Dx
: location of the lesion at diagnosis.stage_at_Dx
: stage of the lesion at diagnosis.is_nodule
: whether the image was labeled as positive for nodule.is_malignant_nodule
: 1 if is_nodule ==1 and cancer_pathology ==1etc
: miscellaneous notes about the image.
Hyunsuk Yoo, Medical Director at Lunit Inc. - [email protected]
Hyunsuk Yoo is an employee of Lunit.