Auto Weight Dilated Convolutional Ensemble Network for the Severity Analysis of Lung Cancer in CT Images

Abstract

Lung cancer is one of the world's leading causes of morbidity and mortality; improving patient outcomes requires an early and precise diagnosis. Lesion and tumor segmentation remains a challenging task in CT images due to their inherent imaging limitations, such as the small size of nodules, heterogeneous textures, blurry boundaries, and adjacent structures, leading to misclassification and difficulty in delineating boundaries. To analyse the severity of lung cancer in CT images, the Auto Weight Dilated Convolutional Ensemble Network (AWDCE-Net) was developed in this article. To extract features of multi-scale lung pulmonary nodules, we created the AD-Net, or auto-weight dilated convolution network. In particular, multi-scale convolutional feature maps were employed by the Auto-weight Dilated convolutional (AD) unit to collect the MA features' auto-weight scales. Using a learnable set of parameters, the AD unit fused convolutional feature maps in encoding layers. The AD unit is a helpful design for feature extraction during the encoding process. We combined the advantages of the U-Net network for both shallow and deep features with the AD unit. AWDCE-Net's exceptional effectiveness in processing lung cancer CT images is demonstrated by experimental evaluation on the IQ-OTH/NCCD dataset, which yielded an accuracy of 99.12% and an F1-measure of 99.12%. With accuracy and F1-score improvements of 2.18 and 1.51%, respectively, these measurements show a significant improvement over popular models.