Lung cancer Screening

Purpose	Lung cancer screening
Tag(s)	Lung cancer, screening, CT, lung nodule
Panel	Thoracic Panel
Define-AI ID	22080015
Originator	Scott J. Adams, MD
Lead	Scott J. Adams, MD
Panel Chair	Carol Wu, MD
Panel Reviewers	Thoracic Panel
License	Creative Commons 4.0
Status	Public Comment
RadElement Set	RDES212

Clinical Implementation

Value Proposition

Multi-center trials have demonstrated that lung cancer screening with low-dose CT imaging in high-risk individuals reduces mortality.1,2 Lung cancer screening is based on the principle that early detection of lung cancer—when cancers appear as lung nodules on CT—allows for timely diagnosis and treatment and better patient outcomes. The primary tasks associated with interpretation of low-dose CT studies for lung cancer screening include detection of lung nodules, measurement of lung nodules, comparison of lung nodules to prior studies (when available), and making a management recommendation, commonly based on the American College of Radiology’s Lung CT Screening Reporting and Data System (Lung-RADS).3 Considering the large number of CT lung cancer screening exams performed at many centers, the time associated with each of these tasks for radiologists, and considering the large proportion of false positives that result from current approaches,4 there is significant potential for artificial intelligence to aid in each of these steps.

Narrative(s)

A 66-year-old male with a 38-pack-year smoking history enters a lung cancer screening program and has a baseline low-dose CT study. An AI algorithm detects all lung nodules on the CT images and highlights all lung nodules for the radiologist to easily review. The AI algorithm provides measurements of all lung nodules, allowing for more reliable comparison between studies. The radiologist validates the lung nodules which have been detected by the AI algorithm, and all lung nodules are automatically described in a draft report in the radiology reporting/dictation system. In addition, the AI algorithm determines the probability of malignancy for each lung nodule. With an AI algorithm which has sufficient sensitivity and specificity, a lung nodule that might otherwise require follow-up based on current guidelines may be determined by the AI algorithm to have a very low risk of malignancy, thereby eliminating the need for interval follow-up and reducing screening costs and patient anxiety. Alternatively, a small lung nodule that might require no follow-up based on current guidelines may be determined by the AI algorithm to have a very high risk of malignancy, leading the patient to have short-term follow-up and an earlier diagnosis of lung cancer.

Workflow Description

CT images are sent to PACS and the AI algorithm. Images are analyzed by the AI algorithm and all candidate pulmonary nodules are segmented and annotated. The radiologist selects candidate pulmonary nodules which are false positives and should be removed from the annotations. Nodule location (lobe and image and series numbers), nodule size (long and short axis measurements and lung nodule volume), nodule attenuation, nodule margins, and nodule shape are determined by the AI algorithm for each lung nodule and validated by the radiologist. The radiologist adds nodules missed by the AI algorithm (false negatives), and re-measures nodules with segmentation/measurement errors.

In cases where prior CT exams are available, the AI algorithm identifies correlates of each lung nodule on prior exams and indicates how the nodule has changed in size.

The AI algorithm determines the probability of malignancy for each lung nodule based on current imaging features, how the nodule has changed in size over time, and clinical information (where available).

All nodule features are summarized and sent to PACS as a secondary DICOM image. Summarized information includes nodule location (lobe and image and series numbers), nodule size (long and short axis measurements and lung nodule volume), nodule attenuation, nodule margins, nodule shape, and malignancy probability score. A description of each lung nodule is pre-populated in a template report in reporting/dictation software.

Based on the probability of malignancy, the AI algorithm provides a suggested Lung-RADS category, which the radiologist can confirm or replace with another category.

Considerations for Dataset Development

Procedure(s):	Low-dose chest CT w/o intravenous contrast
Slice:	1.5 mm or less
Scanner hardware:	Range of vendors and CT scanner models (alternatively, limited to the CT scanner models for which the algorithm will be subsequently used)
Reconstruction kernel:	Lung reconstruction kernels across vendors (alternatively, limited to the lung reconstruction kernels for which the algorithm will be subsequently used). See “QIBA Profile: Lung Nodule Volume Assessment and Monitoring in Low Dose CT Screening”5
Sex at birth:	Male, female
Age:	50 to 80 years (based on United States Preventive Services Task Force recommendations for lung cancer screening6)
Smoking history:	20 pack-year smoking history and currently smoke or have quit within the past 15 years (based on United States Preventive Services Task Force recommendations for lung cancer screening6)
Comorbidities:	Range of background lung disease (COPD, ILD, bronchiolitis, etc.) which may confound lung cancer screening or be lung nodule mimics
Geographic regions:	Various geographic regions (e.g. United States Midwest, West) with varying prevalence of lung nodules secondary to endemic fungal disease, for example
Presence of lung nodules:	CT exams with and without lung nodules (≥3 mm)
Nodule size:	≥3 mm (long axis). Ground truth for long and short axis measurements established by manual or semi-automated methods. Ground truth for volumetric measurements determined by manual segmentation or established semi-automated methods.
Nodule attenuation:	Solid, part-solid, groundglass; see below for additional descriptors of attenuation which may suggest a specific diagnosis such as a granuloma or hamartoma
Nodule shape:	2D: round, oval, triangular, lobular, irregular; 3D: spherical, flat
Nodule margin:	Smooth, irregular, spiculated, ill-defined
Nodule location:	Subpleural, peri-fissural, parenchymal, intraluminal
Nodule diagnosis:	Malignant, benign, or unknown Note: Pathologic diagnosis by core needle biopsy, surgical biopsy, or fine needle aspiration is considered the gold standard for establishing whether a nodule is benign or malignant. If a pathologic diagnosis is not available, benign status can be inferred based on stability over time (suggested time intervals to confer a nodule is benign: stability for 2 years for a solid nodule or 5 years for a ground glass nodule). Malignant status may be inferred based on clinical (thoracic surgery/pulmonology) management of a presumed lung cancer which is not biopsied.

Technical Specifications

Inputs

DICOM Study

Procedure	CT
Views	axial
Data Type	DICOM
Modality	CT
Body Region	chest
Anatomic Focus	lung

Secondary Inputs

Smoking history (pack-years)

Definition	Estimated pack-year smoking history (average number of packs of cigarettes smoked per day multiplied by number of years of smoking)
Data Type	Numeric
Value Set	N/A
Units	Pack-years

Smoking status

Definition	Current smoking status defined as current smoker (an adult who has smoked 100 cigarettes in his or her lifetime and who currently smokes cigarettes), former smoker (an adult who has smoked at least 100 cigarettes in his or her lifetime but who has quit smoking), and never smoker (an adult who has never smoked, or who has smoked less than 100 cigarettes in his or her lifetime)
Data Type	Categorical
Value Set	current smoker former smoker never smoker
Units	N/A

Primary Outputs

Detection of nodule

RadElement ID	RDE612
Definition	The definition of pulmonary nodule detection includes (1) the center x, y, and z coordinate of a candidate nodule bounding box with reference to the superior, anterior, and right-most pixel in the volume (referencing the patient for sidedness, zero indexed) and (2) the dimensions of a bounding box in pixels (x, y, z).
Data Type	Categorical
Value Set	Indeterminate absent present
Units	N/A

Nodule attenuation

RadElement ID	RDE613
Definition	Determine mean and range of HU density of nodule
Data Type	Categorical
Value Set	fat density groundglass part solid - solid component <50% of nodule part solid - solid component ≥50% of nodule solid additional features: calcification cavitation cystic lucencies air bronchograms
Units	N/A

Nodule diameter

RadElement ID	RDE607
Definition	Measure the long and short axis diameters of each nodule. For nodules less than 10 mm, summarize dimensions with an average of the maximal long and short axis measurements. For part solid nodules, return both the overall measurements and measurements of the solid component. Report to nearest 0.1 mm.
Data Type	Numeric
Value Set	N/A
Units	mm

Nodule volume

RadElement ID	RDE608
Definition	Measure the volume of each nodule.For part solid nodules, return both the overall and solid component volumes.
Data Type	Numeric
Value Set	N/A
Units	mm3

Secondary Outputs

Nodule shape

RadElement ID	RDE614
Definition	Classify shape of nodule
Data Type	Categorical
Value Set	indeterminate round oval triangular irregular lobular other
Units	N/A

Nodule margin

RadElement ID	RDE615
Definition	Classify shape/margin of nodule
Data Type	Categorical
Value Set	indeterminate smooth lobulated spiculated ill defined
Units	N/A

Nodule location

RadElement ID	RDE616
Definition	State in which pulmonary lobe the nodule is located
Data Type	Categorical
Value Set	indeterminate right upper lobe right middle lobe right lower lobe left upper lobe left lower lobe fissural in right lung fissural in left lung intraluminal/endobronchial
Units	N/A

Nodule change (categorical)

RadElement ID	RDE1389
Definition	Change in diameter, or volume, or attenuation over time if prior CT studies are available
Data Type	Categorical
Value Set	indeterminate unchanged (to the nearest 1 mm) increase in size decrease in size decreased attenuation increased attenuation
Units	N/A

Nodule change (numeric)

RadElement ID	RDE1390
Definition	Change in diameter and/or volume of nodule over time if prior CT studies are available
Data Type	Numeric
Value Set	N/A
Units	mm and/or mm3

Probability of malignancy

RadElement ID	RDE609
Definition	Probability the nodule is malignant
Data Type	Numeric
Value Set	[0,1]
Units	N/A

Note: exams from a single patient over different time points should be linked by a common identifier with a variable indicating the actual or relative date of each study.

Future Development Ideas

There are various ways the probability of malignancy may be presented and clinically used by radiologists, clinicians, and patients. Further research is required to define thresholds which lead to specific management recommendations and determine how malignancy probability scores should be used in combination with Lung-RADS.7 Incorporating rate of nodule growth into risk prediction strategies is another area for further development. Incorporating biomarkers and additional clinical risk factors for lung cancer as inputs for AI algorithms may further improve the accuracy of malignancy prediction. Follow-up intervals may be personalized based on the probability of malignancy as determined by the AI algorithm in combination with other patient risk factors.
There is potential for integration with other AI algorithms for detection of other lung and non-lung pathology, as well as AI algorithms for non-interpretive tasks, such as identifying patients eligible for lung cancer screening and ensuring patients are appropriately scheduled for follow-up investigations.

References

1. National Lung Screening Trial Research Team, Aberle DR, Adams AM, et al. Reduced lung-cancer mortality with low-dose computed tomographic screening. N Engl J Med. 2011;365(5):395-409. doi:10.1056/NEJMoa1102873.

2. De Koning HJ, Van Der Aalst CM, De Jong PA, et al. Reduced lung-cancer mortality with volume CT screening in a randomized trial. N Engl J Med. 2020;382(6):503-513. doi:10.1056/NEJMoa1911793.

3. American College of Radiology. Lung CT screening reporting and data system (Lung-RADS). 2019. https://www.acr.org/Clinical-Resources/Reporting-and-Data-Systems/Lung-Rads. Accessed March 24, 2021.

4. Pinsky PF, Gierada DS, Black W, et al. Performance of Lung-RADS in the National Lung Screening Trial. Ann Intern Med. 2015;162(7):485. doi:10.7326/M14-2086.

5. CT Volumetry Technical Committee. Lung Nodule Assessment in CT Screening Profile - 2017, Quantitative Imaging Biomarkers Alliance. Publicly Reviewed Draft. Available at: https://qibawiki.rsna.org/images/f/fb/QIBA_CT_Vol_LungNoduleAssessmentInCTScreening_2017.07.rev15.pdf.

6. US Preventive Services Task Force, Krist AH, Davidson KW, et al. Screening for Lung Cancer: US Preventive Services Task Force Recommendation Statement. JAMA. 2021;325(10):962-970. doi:10.1001/jama.2021.1117.

7. Adams SJ, Mondal P, Penz E, Tyan C-C, Lim H, Babyn P. Development and cost analysis of a lung nodule management strategy combining artificial intelligence and Lung-RADS for baseline lung cancer screening. J Am Coll Radiol. 2021:1-11. doi:10.1016/j.jacr.2020.11.014.