Evo-GUNet3++: Using evolutionary algorithms to train UNet-based architectures for efficient 3D lung cancer detection

Abstract

The early detection of malignant lung nodules can strongly increase the chances of life in lung cancer patients. A computer tomography scan represents an effective way to identify and locate malignant nodules in the body and monitor their growth. However, the reading and interpretation of tomography scans are subject to errors that can be reduced with a second reader. The adoption of image processing systems can reduce the possibility of errors and can support radiologists in ensuring multiple readings of tomography scans. This study proposes a new approach for accurate 3D lung nodule detection starting from computer tomography scans. This work exploits an evolutionary algorithm to build variants of a UNet-based architecture, called GUNet3++, to detect patients affected by lung cancer, from the analysis of CT-scan images of lungs. The approach is validated on the LIDC-IDRI real dataset and results show that it improves segmentation quality metrics (IoU and Dice) over baselines, leading to better 3D models reconstruction of lesions.