Predicting non-visible future tumour from baseline low dose CT using deep learned features

Sheng Yu, Yan Wen, Carolyn Horst, Balaji Ganeshan, Spencer Angus Thomas, Reyer Zwiggelaar, Richard Lee, Xujiong Ye, Matthew Blackledge

Keywords: Lung cancer screening, Tumour prediction, Image registration

Abstract: Studies have shown that yearly screening with low-dose computed tomography (LDCT) effectively reduces lung cancer mortality (NLST Research Team, 2011b). With the increas- ing number of deep learning tools that are trained on large collection of scans it has been possible to automatically report lung nodules (Venkadesh et al., 2023). We hypothesized that deep learning might also be able to predict the risk of future malignancies based on LDCT imaging in which no tumours are presently visible. This would provide a triage mechanism for identifying patients who would benefit from yearly screening versus those who might attend biannual screening. We use data from The National Lung Screen Trial (NLST) (NLST Research Team, 2011a) and compare the accuracy of multiple pre-trained classification models from the Keras library to predict a slice-by-slice risk of future tumour occurrence. The best performing model can achieved an AUC of 0.741 demonstrating a successful classifier.

Submission Number: 131