Convolutional Neural Networks Detect Local Infiltration of Lung Cancer Primary Lesions on Baseline FDG-PET/CT

Margarita Kirienko, Martina Sollini, Giorgia Silvestri, Serena Mognetti, Emanuele Voulaz, Lidija Antunovic, Alexia Rossi, Luca Antiga and Arturo Chiti

Abstract: The aim of our work is to develop an algorithm for the classification of lung cancer as T1–T2 or T3–T4 on staging fluorodeoxyglucose positron emission tomography (FDG-PET)/CT images. We retrospectively selected a cohort of 632 patients from patients submitted to FDG-PET/CT, over a 6-year period, for the purpose of staging a suspected lung lesion, within 60 days before biopsy or surgical procedure. Post-acquisition processing was performed to generate an adequate dataset for the convolution neural network (CNN) analyses. The input of CNNs in this study was a bounding box on both PET and CT images, cropped around the lesion centre identified by two nuclear medicine physicians. The algorithm developed and tested in the present work achieved an accuracy of 83.9%, 76.2% and 69.8% in the training, validation and test set, respectively, for the identification of T1–T2 and T3–T4 lung cancer. We obtained proof of concept that CNNs can be used as a tool to assist in the staging of patients affected by NSCLC. Further research in this field is already being addressed by our group.

Keywords: lung cancer, convolutional neural networks, staging, PET/CT, artificial intelligence, deep learning