2023 World Conference on Lung Cancer (Posters)-P1.20. Computed Tomography-based Deep Learning Model for Spread Through Air Spaces Prediction in Ground-Glass Predominant Lung Adenocarcinoma

P1.20. Computed Tomography-based Deep Learning Model for Spread Through Air Spaces Prediction in Ground-Glass Predominant Lung Adenocarcinoma - PDF(Slides)

Back to course

Pdf Summary

Researchers from National Taiwan University Hospital have developed a deep learning model called STAS-DL (Spread Through Air Spaces Deep Learning) that accurately predicts tumor spread through air spaces (STAS) in ground-glass-predominant early-stage lung adenocarcinoma. Predicting STAS preoperatively is crucial for effective surgical planning in these cases, but no previous deep learning models have been developed for this purpose. The study included 581 patients from two institutes between 2015 and 2019, with 458 patients in the training set and 123 in the testing set. <br /><br />The STAS-DL model outperformed other methods, achieving an AUC of 0.82 and 74% accuracy in the testing set. This was significantly better than the STAS-DLwoSCG model, which achieved an accuracy of 70%, and superior to the predictions made by five physicians with an AUC of 0.68. The STAS-DL model also achieved the highest standardized net benefit when compared to other methods and physicians. <br /><br />The STAS-DL model utilizes a solid components gated (SCG) feature extraction technique to predict STAS. The model was developed and validated with a ground-glass-predominant nodule and underwent preprocessing steps including volume of interest extraction, isotropic interpolation, gray-level normalization, and segmentation. The training was initially performed using a training cohort, with a 3-fold cross-validation for internal validation. External validation was then performed using a testing cohort and compared against predictions made by physicians. <br /><br />Overall, the researchers concluded that the STAS-DL model has great potential for accurately predicting STAS preoperatively in patients with ground-glass-predominant early-stage lung adenocarcinoma. The model may assist clinicians in making informed decisions regarding surgical planning for these patients.

Asset Subtitle

Mong-Wei Lin

Meta Tag

Speaker Mong-Wei Lin

Topic Pathology & Biomarkers: Artificial Intelligence in Pathology

Keywords

National Taiwan University Hospital

deep learning model

STAS-DL

tumor spread through air spaces

lung adenocarcinoma

surgical planning

ground-glass-predominant

AUC

physicians

preoperative prediction