Deep learning and classical segmentation techniques for lung cancer imaging

Abstract

Lung cancer remains one of the most lethal malignancies worldwide, with prognosis heavily reliant on early and accurate diagnosis. The segmentation of medical images is pivotal in lung cancer detection, offering precise delineation of tumors for effective clinical decision-making. Traditionally, lung segmentation has employed classical image-processing methodologies, including thresholding, clustering, and edge detection. These techniques, however, have limitations, such as susceptibility to variability and manual intervention requirements, leading to inconsistent outcomes. Recently, deep learning (DL) architectures have revolutionized medical image segmentation, demonstrating superior performance through automated, high-dimensional feature extraction capabilities. In particular, convolutional neural networks (CNNs) such as U-Net and its advanced variants, like U-Net++, have significantly improved segmentation accuracy. The integration of dense skip connections and deep supervision within U-Net++ bridges the semantic gap between the encoder and decoder paths, leading to enhanced prediction accuracy. Moreover, the U-Net Transformer (UNETR) architecture, leveraging transformers alongside the traditional U-Net structure, provides remarkable results due to its ability to capture global contextual information from 3D medical images. The application of DL approaches extends further into the incorporation of hybrid models, including self-supervised neural networks, to classify segmented lung images into benign or malignant categories. Such integration offers an end-to-end diagnostic system, reducing errors and variability introduced by manual segmentation methods. Through comprehensive evaluation on publicly available datasets such as Decathlon And Montgomery–Shenzhen, DL techniques have consistently outperformed traditional methods, achieving segmentation accuracies surpassing 98%. This chapter critically reviews classical and advanced deep segmentation techniques, evaluates their efficacy in clinical scenarios, discusses the advantages and limitations of each method, and outlines future directions for improved diagnostic systems in lung cancer imaging.