Incidental Pulmonary Nodules on Chest Radiograph

Purpose

Detection and characterization of incidental pulmonary nodules on chest radiographs (CXR). These are nodules that are detected on CXRs performed for other reasons than lung cancer screening.

Tag(s)

 

Panel

Thoracic

Define-AI ID

08190004

Originator

Thoracic Panel

Panel Chair

Warren B. Gefter, MD; Eric J. Stern, MD

Panel Reviewers

Thoracic Panel

License

Creative Commons 4.0

Status

Released for Comments

Clinical Implementation


Value Proposition

Lung cancer, the leading cause of cancer-related deaths in both women and men, frequently presents as a pulmonary nodule on chest radiographs (CXRs) or CT scans. While low-dose CT is utilized for lung cancer screening, chest radiography, being among the most highly utilized diagnostic imaging procedures worldwide, is the most common thoracic imaging study in which incidental lung cancers are discovered. Nonetheless, interpretation of chest radiographs is challenging and prone to many reading errors. Thus nodules are frequently missed on CXRs, with studies showing approximately 20-30% (even up to 90%) seen only in retrospect. The causes for these frequent errors are multifactorial, including: overlapping anatomic structures such as the ribs, clavicles, thoracic spine, pulmonary vessels, heart, mediastinum and diaphragms; errors in visual search, lesion recognition or decision-making; and suboptimal image quality. Small, ill-defined nodules with low attenuation and conspicuity are particularly susceptible to being overlooked. As early detection of lung cancer reduces mortality, missed or delayed diagnosis due to these CXR errors may negatively impact patient survival.

Furthermore, such errors carry significant medicolegal risks, being the second most common cause (after breast cancer) for malpractice litigation in radiology. Algorithms based upon machine learning therefore offer an important use case to assist radiologists in more accurate detection, characterization, and any communication and recommendation for further study of these nodules. This may be particularly true for less experienced radiologists or in places without access to radiology expertise. These algorithms show promise in improving upon traditional CAD (computer-assisted detection) systems.

Narrative(s)

A 45-year-old man with cough and fever has a CXR for evaluation of possible pneumonia. Algorithm evaluates the lungs and detects a non-calcified, irregular-shaped nodule at the right lung apex partially obscured by the anterior first rib. Lesion is highlighted on annotated image, so as not to be overlooked by the radiologist. Radiologist confirms that this is a new finding compared with older CXRs and recommends further evaluation with a chest CT scan. Appropriate communication with the referring clinician is made.

A 40-year-old woman, never-smoker, undergoes a CXR for a newly positive PPD. Algorithm evaluates the lungs and detects a subtle smooth, round, non-calcified nodule in the retrocardiac region of the left lung. Lesion is highlighted on image annotated by algorithm. No prior CXRs are available. Radiologist recommends further evaluation with a chest CT. Appropriate communication with the referring physician is made.

A 69-year-old man with a long history of cigarette smoking has a CXR to evaluate a chronic cough. Algorithm evaluates the lungs and detects a subtle, juxtavascular nodule adjacent to the right hilum. The lesion is highlighted on an image annotated by the algorithm so as not to be missed by the radiologist. In retrospect the lesion can be seen on an earlier CXR and has enlarged in the interim. The radiologist recommends further evaluation with a chest CT and assures appropriate communication with the referring physician.

Workflow Description

CXR sent to PACS and the AI engine. Image analyzed by AI algorithm, which detects and characterizes the incidental pulmonary nodule(s). Annotated image(s) highlighting each incidental pulmonary nodule with associated nodule characteristics sent to PACS. Icon indicating AI nodule detection may appear on worklist for prioritization.

Considerations for Dataset Development


Procedures(s): {CXR, CR, DR, dual-energy, and bone-suppression CXRs}

View(s): { PA, lateral, AP, apical lordotic, obliques }

Sex at Birth: {Male, Female}

Age: [18, no limit]

Nodule Validation: { CT within 1 month of CXR. Corresponding nodule location on CXR confirmed by chest radiologists. }

Nodule Characteristics: { Attenuation based upon CT confirmation (solid, part-solid, groundglass, internal fat density, calcification, cavitary);
Size (5 mm to 40mm)
Shape (round, oval, triangular, lobular, irregular);
Margin (smooth, irregular, spiculated);
Location (broad sampling of lung regions, apex to base, central to peripheral) } 

Comorbidities: { Smokers, non-smokers, COPD, travel/exposure history, other primary malignancy or history of primary malignancy, bronchitis, bronchiolitis, pneumonia, tuberculosis, fungal and other pulmonary infections, focal inflammatory lesions, usual interstitial pneumonia and other diffuse lung diseases, pleural effusion. }

Other: { Range of CXR technologies (CR, DR, dual-energy, bone suppression) and patient population demographics. Include normal CXRs without nodules, as well as those with single and multiple nodules. Range of nodule conspicuity. Datasets should be enriched with more challenging nodules prone to human error, including small lesions < 1 cm; lesions located in the apices/upper lobes, retrocardiac, perihilar and retrophrenic areas; and nodules with low conspicuity }

Technical Specifications


Inputs

DICOM Study

Procedure

XRAY, Chest

Views

CXR: PA, lateral, AP, apical lordotic, obliques

CR, DR, dual-energy, and bone-suppression CXRs

Data Type

DICOM

Modality

XRAY

Body Region

Chest

Anatomic Focus

Lung

Pharmaceutical

N/A

Scenario

N/A

 

Primary Outputs

Nodule Detection

RadElement ID

 

Definition

The center x and y coordinates of a candidate nodule bounding box with reference to the superior and right-most pixel in the bounded area (referencing the patient for sidedness, zero indexed)

Data Type

Array

Value Set

 

Units

N/A

 

Nodule Localization

RadElement ID

 

Definition

The dimensions of a bounding box in pixels (x and y)

Data Type

Array

Value Set

 

Units

N/A

 

True Lung Nodule Probability

RadElement ID

 

Definition

The probability that the bounded CXR opacity represents a true lung nodule.

Data Type

Numerical

Value Set

[0,1]

0-Opacity definitely not a lung nodule

1-Opacity definitely a lung nodule

Units

N/A

 

Nodule Attenuation

RadElement ID

 

Definition

Determine density of a nodule

Data Type

Categorical

Value Set

0 – fat density

1 – groundglass

2 – part solid

3 – solid

4 – calcification

5 – cavitation

6 – cystic lucencies

7- air bronchograms

Units

N/A

 

Nodule Size

RadElement ID

 

Definition

Measure diameters of nodules

Data Type

Numerical

Value Set

 

Units

mm

 

Nodule Shape

RadElement ID

 

Definition

Describe nodule shape

Data Type

Categorical

Value Set

1-round

2-oval

3-triangular

4-irregular

5- lobular

Units

N/A

 

Nodule Margin

RadElement ID

 

Definition

Describe the shape/margin of nodule

Data Type

Categorical

Value Set

1-smooth

2-lobulated

3-spiculated

Units

N/A

 

Nodule Location

RadElement ID

 

Definition

State lung region in which nodule is located

Data Type

Categorical

Value Set

1-right lung, upper third

2-right lung, middle third

3-right lung, lower third

4-left lung, upper third

5-left lung, middle third

6-left lung, lower third

Units

N/A

 

Secondary Outputs

Nodule Growth

RadElement ID

 

Definition

Change in nodule size (longest diameter over time if older CXRs available)

Data Type

Categorical

Value Set

0-Unknown

1-Stable

2-Growth

3-New

4-Shrinkage

5-Resolved

Units

N/A

 

Probability of Malignancy

RadElement ID

 

Definition

Likelihood the nodule is malignant based on nodule characteristics, patient demographics and smoking history

Data Type

Numerical

Value Set

[0,1]
0-Benign

1-Malignant

Units

N/A

Future Development Ideas


Tracking of follow-up imaging

Automated tracking of whether recommended additional imaging studies have been ordered, scheduled, and performed and reminders to ordering clinicians or patients

Probability of Malignancy

Algorithm may output probability of malignancy of detected nodule(s) based upon CXR imaging features, smoking history and other clinical parameters.