2023 North America Conference on Lung Cancer (NACLC) - Posters and Abstracts-PP01.49 (Poster) END TO END MODEL FOR NODULE DETECTION AND CHARACTERIZATION IN LUNG CANCER SCREENING: PERFORMANCES AND SUBPOPULATION ANALYSIS

PP01.49 (Poster) END TO END MODEL FOR NODULE DETECTION AND CHARACTERIZATION IN LUNG CANCER SCREENING: PERFORMANCES AND SUBPOPULATION ANALYSIS

Back to course

Pdf Summary

Researchers from Median Technologies in France have developed an end-to-end model for the detection and characterization of lung nodules in lung cancer screening. The model consists of two parts: 3D-CNN detection models to locate nodules, and 3D-CNN models to determine the malignancy risk of detected nodules. The model was trained using data from the National Lung Screening Trial (NLST) and Lung Image Database Consortium (LIDC), and was evaluated on a test set of 2163 NLST patients.<br /><br />The model demonstrated robust performance in detecting and predicting the malignancy risk of nodules in lung cancer screening populations. It outperformed the NLST Brock model, which relies on clinician's nodule detection and feature assessment. The model's performance was consistent across different characteristics, indicating its potential to improve patient management in screening programs and early cancer diagnosis.<br /><br />The researchers also compared the model's performance to the NLST Brock model on various subsets of nodules based on size, spiculation, and solidity. The model consistently outperformed the NLST Brock model across these subsets, demonstrating its efficacy in differentiating between malignant and benign nodules.<br /><br />The model achieved an area under the ROC curve (AUC-ROC) of 0.987 at the nodule level on the localized subset of nodules, compared to the NLST Brock model's AUC-ROC of 0.971. When stratified by size, the AUC-ROC was highest for nodules 4-10mm and 10-20mm, and lower for nodules 20-30mm. The model also performed well in distinguishing between spiculated and non-spiculated nodules, as well as solid and non-solid nodules.<br /><br />Overall, the end-to-end model demonstrated strong performance in detecting and characterizing lung nodules, and has the potential to enhance lung cancer screening and patient management.

Asset Subtitle

Charles Voyton

Keywords

lung nodules

lung cancer screening

end-to-end model

3D-CNN detection

3D-CNN models

malignancy risk

NLST

LIDC

patient management

early cancer diagnosis