Abstract 1190P
Background
TNM staging is used to predict outcome and guide management in patients with lung cancer, but there are known discrepancies between clinical and pathological staging. A computational approach to staging may improve both the accuracy and consistency of clinical staging. This study examines a machine learning method to predict pathological T-stage and compares its performance to that of clinical staging and nodule diameter.
Methods
The US National Lung Screening Trial dataset was curated, such that each reported nodule was located, measured and diagnostically characterised. A subset for this study was built by selecting all solid and semi-solid malignant pulmonary nodules greater than 6mm in diameter. Given dataset constraints, only T1a, T1b, T2a and T2b patients were classified. A Convolutional Neural Network (CNN) was trained on CT images from 415 lung cancer patients to classify cancers as either T1 (incl. T1a/T1b) or T2 (incl. T2a T2b). This was derived from a previously trained CNN designed to classify benign-vs-malignant nodules. A 10-fold cross-validation strategy was used, where, in each fold, 8 (of the 10 equally-sized) subsets were used for training, one for model tuning, and one for validation. This was repeated 10 times, each time using a different subset for validation. Performance was evaluated using the Area-Under-the-ROC-Curve (AUC), stage-wise accuracy and balanced accuracy, using pathological T-stage as the ground truth. Clinical and pathological T-stage data were taken from the NLST records.
Results
330 T1 cases and 85 T2 cases were selected. We found that clinical staging balanced accuracy is not significantly better than that of the model (P=0.602). Table: 1190P
Classification results for each pathological T-stage predictor. All values are given in % with 95% confidence interval bounds in parentheses. Balanced accuracy accounts for the T1/T2 case number imbalance by weighting each stage accuracy equally
| Predictors | AUC | T1 Accuracy | T2 Accuracy | Balanced Accuracy |
| CNN | 79.6 (72.7, 86.1) | 90.0 (85.8, 93.9) | 50.8 (38.0, 63.8) | 70.4 (63.6, 77.1) |
| Diameter (remeasured) | 74.4 (65.7, 82.8) | 98.1 (96.1, 99.5) | 39.0 (26.8, 51.8) | 68.5 (62.2, 75.0) |
| Clinical Staging (from NLST records) | - | 100.0 (100.0, 100.0) | 39.0 (26.9, 51.6) | 69.5 (63.3, 75.8) |
Conclusions
Based on this analysis of T1 and T2 lung cancers, CNN determination of T1/2-stage, a critical factor in the management of lung-cancer patients, has equivalent accuracy to clinical T-staging. With further development, this algorithm could prove a useful tool to aid physician workflow and to increase the consistency of clinical T-staging.
Clinical trial identification
Editorial acknowledgement
F. Hoffmann-La Roche was given the opportunity to review and approve the abstract for scientific accuracy only.
Legal entity responsible for the study
Timor Kadir.
Funding
F. Hoffmann-La Roche.
Disclosure
N. Waterfield Price: Shareholder/Stockholder/Stock options, Full/Part-time employment: Optellum Ltd. T. Kadir: Leadership role, Shareholder/Stockholder/Stock options, Full/Part-time employment, Officer/Board of Directors: Optellum Ltd. P. Massion: Advisory/Consultancy, Advisory board (non-funded): Optellum Ltd. F. Gleeson: Advisory/Consultancy, Shareholder/Stockholder/Stock options, Advisory board (non-funded)(non-funded): Optellum Ltd. L. Freitag: Shareholder/Stockholder/Stock options, Full/Part-time employment: Optellum Ltd.