Summary
Excerpt |
---|
Automated or semi-automated algorithms intended for chest CT analyses typically require the creation of a 3D map of the thoracic volume as their initial step. Identifying this anatomic region precedes fundamental tasks such as lung structure segmentation, lesion detection, and radiomics feature extraction in analysis pipelines. However, automatic approaches to segment the thoracic volume maps struggle to perform consistently in subjects with diseased lungs, yet this is exactly the circumstance for which pipeline analyses would be most useful. To address this need, we have created a dataset of thoracic volume segmentations on subjects with diseased lungs. These will help the research community compare and contrast their approaches for this foundational processing step on clinically relevant data.This dataset consists of left and right thoracic volume segmentations delineated on 402 CT scans from The Cancer Imaging Archive NSCLC Radiomics collection. Thoracic segmentations include lung parenchyma, tumor, atelectasis, and effusion when present. On scans where effusion is present, separate segmentations labeling pleural effusion alone are also provided. Thoracic segmentations were generated automatically by a U-Net, manually corrected by a medical student, and revised by a radiation oncologist or a radiologist. Pleural effusion segmentations were manually delineated by a medical student and revised by a radiologist. Expert GTV segmentations already provided by the NSCLC Radiomics collection helped inform our segmentations. The thoracic segmentations were developed to study image feature symmetry between left and right thorax anatomy as a predictor for tumor localization. Previous studies have developed novel deep learning architectures that leverage deviations from natural anatomic symmetry to predict endpoints. For example, a deep learning architecture called DeepSymNet localized stroke cores in brains by comparing information acquired from convolutional neural networks that analyzed brain hemispheres. Researchers interested in developing novel deep learning architectures that predict NSCLC endpoints using a similar methodology may find our thoracic segmentations useful. Researchers interested in discriminating between tumor and effusion using imaging biomarker inputs may find our pleural effusion segmentations useful, especially when paired with the GTV segmentations provided in the NSCLC Radiomics collection. This dataset was utilized in a clinical informatics study correlating autosegmentation accuracy as gauged by several common metrics with time required for manual corrections. Manual correction time, accuracy metrics, and other tabular data are provided and extend the tabular data accompanying the NSCLC Radiomics CT datasets. |
Acknowledgements
We would like to acknowledge the individuals and institutions that have provided data for this collection:
...
Localtab Group | |||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||
---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
|
...