Pulmonary Nodule Classification

About
Interfaces
This algorithm implements all of the following input-output combinations:
Inputs | Outputs | |
---|---|---|
1 |
|
|
Model Facts
Summary
This algorithm predicts the pulmonary nodule type using deep learning algorithms. It uses the same ensemble of convolutional neural networks (ensemble of multi-view ResNet50 and Inception-v1 3D) employed by Venkadesh and colleagues (2021) for pulmonary nodule malignancy risk prediction. It utilizes the same training and external validation datasets as Ciompi and colleagues (2017).
Mechanism
Outcome | Pulmonary nodule type classification (solid, part-solid, non-solid, calcified, perifissural, solid-spiculated) |
Output | Categorical label indicating the predicted nodule type |
Target population | High-risk population eligible for lung cancer screening |
Input data source | Low-dose chest CT examination and coordinate(s) of pulmonary nodule(s) |
Training data location and time-period | Participants undergoing lung cancer screening in the Multicentric Italian Lung Detection Trial (between 2005 and 2014) |
Reference standard | Radiologist consensus annotations based on CT imaging and clinical guidelines |
Model type | Convolutional neural network |
Validation and Performance
Metric | SOTA (Ciompi and colleagues, 2017) | Algorithm |
---|---|---|
Accuracy | 72.8% | 80.3% |
Balanced Accuracy | - | 76.5% |
FSolid | 64.2% | 86.6% |
FCalcified | 88.9% | 90.4% |
FGroundGlassOpacity | 80.0% | 82.6% |
FSemiSolid | 71.7% | 53.3% |
FPerifissural | 77.3% | 76.3% |
FSolidSpiculated | 62.7% | 42.4% |
Table: Evaluation results of the algorithm on the DLCST dataset. The accuracy, balanced accuracy, and the one-versus-rest F-score for each class are reported.
Uses and Directions
For research use only. The algorithm is intended to be used only on pulmonary nodules detected at low-dose chest CT examinations of high-risk participants eligible for lung cancer screening programs.
Benefits: Accurate classification of pulmonary nodule type may improve radiologists’ assessments, aid in risk stratification, and support consistent follow-up recommendations.
Target population and use case: High-risk individuals participating in lung cancer screening undergo annual low-dose CT examinations. Radiologists evaluating chest CT images must characterize pulmonary nodules to inform follow-up recommendations and clinical decision-making.
General use: This model is intended to be used by radiologists for classifying pulmonary nodules detected at screening CT into predefined categories (solid, part-solid, non-solid, calcified, perifissural, or solid-spiculated). The model is not a diagnostic tool for malignancy and is not intended to guide or drive clinical care. It is designed to assist radiologists in characterizing nodules alongside other clinical and imaging findings.
Appropriate decision support: The model classifies pulmonary nodule X as [solid/part-solid/non-solid/calcified/perifissural/solid-spiculated]. The referring radiologist reviews the classification in conjunction with other clinical and imaging data to determine the appropriate follow-up recommendation.
Before using this model: Validate the model retrospectively and prospectively on a representative diagnostic cohort to ensure its performance aligns with the intended clinical setting.
Safety and efficacy evaluation: TBD
Warnings
Risks: Even when used correctly, clinicians relying on this model may misclassify pulmonary nodules. Misclassification could impact follow-up recommendations, potentially delaying the identification of malignant nodules or leading to unnecessary interventions. Patients misclassified as high risk may undergo unwarranted procedures, exposing them to procedural risks and increased healthcare costs.
Inappropriate Settings: This model was not trained on CT examinations from a general clinical or diagnostic setting. It was specifically developed using a high-risk population undergoing lung cancer screening with low-dose CT examinations. Do not apply this model to other populations or settings without rigorous validation.
Clinical rationale: The model does not provide an explanation for its classifications and lacks interpretability. Clinical users should consider the model’s output alongside other imaging and clinical data to make informed decisions rather than relying solely on its predictions.
Inappropriate decision support: The model was designed for pulmonary nodule classification and may not perform reliably outside the intended high-risk screening population. It is not a diagnostic tool and should not be used as the primary determinant for clinical decision-making regarding malignancy or treatment.
Generalizability: The model was primarily trained using pulmonary nodule data from the Multicentric Italian Lung Detection Trial and validated with cases from the Danish Lung Cancer Screening Trial. It should not be deployed in external settings without further validation to ensure reliability.
Discontinue use if: Concerns arise from clinical staff regarding the model’s reliability or applicability to the intended use case, or if significant shifts in imaging protocols, scanner types, or patient demographics occur, necessitating model retraining.
Common Error Messages
Please ignore the warnings. This will be fixed later.
Information on this algorithm has been provided by the Algorithm Editors, following the Model Facts labels guidelines from Sendak, M.P., Gao, M., Brajer, N. et al. Presenting machine learning model information to clinical end users with model facts labels. npj Digit. Med. 3, 41 (2020). 10.1038/s41746-020-0253-3