Computational Imaging and Bioinformatics Laboratory, Harvard Medical School, Boston, MA
Tafadzwa Lawrence Chaunzwa , David C. Christiani , Michael Lanuti , Andrea Shafer , Nancy Diao , Raymond H. Mak , Hugo Aerts
Background: There is a growing body of evidence suggesting radiomic phenotypes can augment prognostic power, when used in combination with clinical features and tumor genomic profiles in lung cancer. In this study we present a deep-learning model that can act as a non-invasive prognostic biomarker in patients with Stage-I Non-Small Cell Lung Cancer (NSCLC). Our model would be able to assign patients to short term or long term survival groups, based on CT characteristics. Methods: Pretreatment CT studies were retrieved for 299 patients who underwent surgery for Stage-I NSCLC at MGH between 2004-2010. Image pre-processing included manual tumor identification, and isotropic rescaling of CT data. Further data curation resulted in a final cohort of 186 patients. Median follow-up from time of diagnosis was 2.9 years and 9.7% of patients were deceased at 2 years. To mitigate bias against a low probability event (mortality), data augmentation was performed yielding 242 50x50 pixel patches to feed into our model. A pre-trained 16 layer deep neural network (VGG-16) was used to perform visual recognition and data analysis. Fine-tuning of the last two convolutional blocks and a fully-connected classifier was performed with a training set of 144 labeled CT scans, matched to one of two groups based on 2 year survival. 34 samples were used for initial cross-validation. Data containing variations of imaging scanners and protocols were used in training to create a model that is robust for the variations. Results: Our model learned to classify patients with long term vs short term survival in an independent validation set of 64 samples with 75% accuracy and AUC = 0.798. In comparison, a multivariate linear regression model of conventional clinical prognostic factors (age, gender, tumor stage, histology, and smoking status) had a lower predictive performance (AUC = 0.665). Event rates were balanced between training and independent-validation groups. Conclusions: These findings suggest that Artificial Intelligence-enhanced radiomic feature extraction and predictive modeling can aid the clinician in assessing the benefits of treatment for patients with early-stage NSCLC.
Disclaimer
This material on this page is ©2024 American Society of Clinical Oncology, all rights reserved. Licensing available upon request. For more information, please contact licensing@asco.org
Abstract Disclosures
2023 ASCO Annual Meeting
First Author: Yan Liu
2023 ASCO Annual Meeting
First Author: Amogh Hiremath
2020 ASCO Virtual Scientific Program
First Author: Erica C. Nakajima
2023 ASCO Annual Meeting
First Author: Yan Liu