Abstract
Introduction
Machine learning algorithms may improve efficiency and accuracy of pathologic response (PR) assessment in surgically resected lung cancers following neoadjuvant therapy. The aim of this study was to develop digital models for quantifying tumor bed (TB) area and residual viable tumor (VT), and to compare these results to previously published assessments of PR by pathologists from the IASLC reproducibility study.Methods
Manual pathologist annotations (N=15,564) of regions including TB and VT were used to train convolutional neural network model (digital AI) and a Convex Hull algorithm (CHA). PR was determined by the percentage of VT in the TB area. The pathologist determined the average PR (APR) across slides (unweighted) which was compared to the weighted average for digital AI and CHA. The concordance between pathologist APR, digital AI and CHA was calculated, and correlated with outcomes.Results
There was a strong correlation between approaches: APR vs. Digital AI (0.97), APR vs. CHA (0.97), and Digital AI vs. CHA (0.99). Digital PR and CHA showed 100% agreement for MPR. The kappa concordance for MPR was 0.82 (95% CI: 0.69, 0.96) for APR versus Digital AI/CHA with 6 discordant cases. The concordance was higher for squamous cell carcinoma (Kappa 0.92, 95% CI: 0.76, 1.0) than for non-squamous carcinoma (Kappa 0.77, 95% CI: 0.59, 0.96). APR and digital AI showed similar relapse-free survival (RFS) and overall survival (OS).Conclusion
The overall high level of agreement supports the utility of the machine learning approaches for evaluations of PR in patients with NSCLC.Authors
- Dacic et al.