Subsolid nodule segmentation
About
Interfaces
This algorithm implements all of the following input-output combinations:
| Inputs | Outputs | |
|---|---|---|
| 1 |
|
|
Model Facts
Summary
This algorithm performs pulmonary nodule segmentation using deep learning-based models. It comprises two dedicated models based on nnU-Net (version 1.7.0): one for segmenting solid nodules and another for subsolid nodules. The solid nodule segmentation model was trained on the LIDC-IDRI dataset (which comprises all types of nodules) and the subsolid nodule segmentation model was trained on the MILD dataset.
Mechanism
- Outcome: Pulmonary nodule segmentation (solid, part-solid, non-solid)
- Output: Segmentation mask outlining the nodule boundaries (0: background, 1: nodule, 2: solid-core)
- Target population: High-risk individuals undergoing lung cancer screening
- Input data source: Low-dose chest CT examination and coordinate(s) of pulmonary nodule(s)
- Solid nodule segmentation: LIDC-IDRI dataset
- Subsolid nodule segmentation: MILD dataset
- Model Type: Convolutional neural network (nnU-Net)
Validation and Performance
| Metric 1 | Solid nodules | Subsolid nodules |
|---|---|---|
| Dice score | 75.0%; published by Venkadesh and colleagues (2023) | 77.6% (non-solid), 76.0% (solid core) |
| Reference standard | Interobserver agreement (LIDC-IDRI) | Pulmonologist utilizing CIRRUS Lung Screening |
Uses and Directions
For research use only. This model is intended to segment pulmonary nodules in low-dose CT scans of high-risk individuals undergoing lung cancer screening.
Benefits: Accurate segmentation of pulmonary nodules provides essential input for malignancy risk assessment, volumetric growth tracking, and standardized reporting in lung cancer screening.
Target population and use case: This model assists radiologists in delineating nodule boundaries for more precise risk stratification and follow-up recommendations in screening settings.
General use: This segmentation model provides automated delineation of solid, part-solid, and non-solid nodules, aiding radiologists in lung nodule evaluation. It is not a diagnostic tool for malignancy determination but serves as a supporting tool for structured radiology workflows.
Appropriate decision support: The model generates a segmentation mask for pulmonary nodule(s) detected in the CT scan. Radiologists can use the measurements from this mask in conjunction with clinical findings to guide nodule characterization and risk assessment.
Before using this model: Retrospective and prospective validation in the intended clinical setting is recommended. Consideration of scanner types, imaging protocols, and patient demographics is essential before deployment.
Safety and efficacy evaluation: TBD
Warnings
Risks: The model may produce inaccurate segmentations. Over-reliance on automated segmentation without human review may lead to errors in malignancy risk assessment and follow-up recommendations.
Inappropriate settings: The model was developed for lung cancer screening populations and has not been validated in diagnostic or incidental nodule detection settings. - It should not be applied to patients not undergoing lung cancer screening without further validation.
Clinical rationale: The model does not provide an explanation for segmentation decisions. Radiologists should review the outputs alongside raw imaging data.
Inappropriate decision support: The model is designed for segmentation only and does not classify nodules as benign or malignant. It should not be used as a standalone diagnostic tool.
Generalizability: The model was trained on datasets from LIDC-IDRI and MILD and validated on DLCST cases. External validation in different populations, imaging protocols, and scanner types is necessary before clinical use.
Discontinue use if: Clinical users report systematic segmentation errors, particularly in nodules with atypical morphologies. Changes in imaging protocols or scanner settings significantly impact segmentation quality, requiring model retraining.
Common Error Messages
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