Enhanced Edge-Deployable Deep Learning System for Lung Cancer Type and Stage Classification Using CT Imagery

One of the most fatal types of cancer worldwide is lung cancer and therefore an accurate and efficient diagnostic system is required for early detection and the stage of the disease. While deep learning approaches have made substantial strides in the diagnosis from Computed Tomography (CT) scans, several current methods are either computationally heavy or depend on ensemble models or transformer-based networks, which are not suitable for healthcare settings with limited computing resources. Furthermore, some procedures only classify cancers and do not include detailed staging using the TNM (Tumor–Node–Metastasis) classification. To solve these problems, this paper presents LiteSOS-Net, a lightweight deep learning framework to classify lung cancer type and predict its TNM stage simultaneously from CT images. The study starts with a thorough review of the literature to define the important gaps in the recent research that would impact clinical usage of the existing methods.

The proposed framework consists of bilinear image resizing, guided-filter denoising, Multi-Peak Generalized Histogram Equalization for enhancing the image contrast, Attention U-Net for segmentation, and a classification network based on ShuffleNet with channel-shuffle gating. The Osprey Optimization algorithm is employed to further optimize the network parameters to increase the accuracy of the network predictions whilst maintaining computational efficiency. Several publicly available benchmark CT image datasets are proposed to evaluate the proposed framework. Its light weight design allows for edge deployment in limited resource healthcare applications. The research gaps identified and methodology proposed is presented in this paper, and the experimental validation and performance analysis in the next phase of the study.