Development of artificial intelligence models to classify pulmonary nodules and improve lung cancer early diagnosis

Abstract

The aims of this thesis were to develop machine-learning models to identify lung nodules, predict the risk of cancer and provide clinical decision support. A structured-query language model was developed at The Royal Marsden Hospital to generate a database of 14,586 patients with lung nodules. Lung (39%), neuro-endocrine (38%) and skin (35%) cancers were most commonly associated with nodules. Nodules patients had more metastatic diagnoses (45% vs 23%, p < 0.001) and a higher mean scan number (6.56 vs 1.93, p < 0.001) at shorter intervals (4.1 vs 5.9 months, p < 0.001). The model was externally validated with high performance (Krippendorf’s Alpha > 0.98). Scans from the LUCADI and LIBRA studies were used to develop small (< 15mm) and large (> 15mm) nodule radiomics predictive vectors (SN and LN-RPV respectively). Features were extracted using TexLab 2.0, and models were developing using LASSO logistic regression. The SN-RPV had an AUC of 0.78 in the test (95% C.I. 0.70-0.86) and external test (95% C.I. 0.71-0.83) sets. For the two-feature LN-RPV, the test set AUC was 0.87 (95% C.I. 0.80-0.93), compared to 0.67 (95% CI 0.55–0.76, DeLong p = 0.002) for the Brock score and 0.83 (95% CI 0.75–0.90, DeLong p = 0.4) for the Herder score. The external test set AUC was 0.75 (95% CI 0.63–0.85). The developed decision-support tool identified 18/22 (82%) malignant nodules in the Herder 10-70% category, and may have led to earlier investigation. Finally, a model was developed to predict nodule spiculation in the LIBRA and NSCLC Radiogenomics studies. The test set AUC for the 7 feature model was 0.90 (95% CI: 0.82- Introduction 5 0.96), and spiculation was associated with worse overall survival (HR 2.0, 95% C.I. 1.00 - 4.01, p = 0.04), the differential expression of 11 genes and suppression of inflammation.