Residual Convolutional Neural Network With Vision Transformer for Lung Nodule Classification

ABSTRACT

Lung cancer is one of the common causes of cancer‐related deaths globally, and improving patient outcomes requires early identification and precise lung nodule categorization. Traditional convolutional neural networks (CNNs) have shown promise in lung nodule classification yet are typically thwarted in their ability to render global contextual information. Recent developments in deep learning, such as vision transformers (VTs) and hybrid architectures, present a potential fix through consideration of both local and global dependencies in medical imaging. A totally innovative novel combination of residual blocks and vision transformers suited for lung nodule classification in the name of Residual Convolutional Network‐Vision Transformer (RCNN‐VT) framework is proposed in this work. By using residual connections that enable efficient gradient flow and using vision transformers for better representation of features, the RCNN‐VT model outperformed previous most sophisticated methods. Experimental results on lung CT datasets significantly outperformed existing CNN‐based and hybrid models. An ablation study further elucidated the contribution of the incorporation of residual blocks and vision transformers, especially for the case of small nodule detection.