A Deep Transfer Learning Approach For Lung Tumour Detection With Resilience Testing Under Suboptimal Conditions

Fatality from lung tumours makes up the most significant proportion of all cancer deaths in the UK, and the 10-year survival rate is less than 10%. This research explores the use of pretrained deep neural networks for lung tumour detection and aims to determine which network is best suited for the task. The report highlights the significance of lung tumours as fatal cancer and the potential of artificial intelligence to improve detection and diagnosis. The methodology involves transfer learning of multiple pretrained neural networks onto a lung tumour dataset and testing robustness against varying levels and types of noise. Implementation and testing have been conducted in Matlab R2022b. Firstly, transfer learning of a number of deep neural networks onto a lung tumour dataset has been performed and a comparison of the performance of all chosen deep neural networks has been obtained. Subsequently, new lung tumour images containing different types of noise were fed through the transfer learned neural networks to determine which networks retain the highest accuracy. The findings suggest that ResNet-50 is the most suitable network for lung tumour detection in both normal and suboptimal conditions.