ResNet-SEMultiRes: A Deep Learning Framework for Primary Lung Cancer Classification Using CT Scan Image

Abstract Lung cancer detection is one of the most challenging tasks in medical image analysis, particularly in computed tomography (CT) images. Classification is critical for accurate diagnosis and appropriate therapy. Diagnosis currently relies on tissue biopsy, an invasive time-consuming procedure that carries potential risks. This study introduces a deep learning-based ResNet-SEMultiRes model, specifically designed to improve lung cancer subtyping from CT scan images. In this study, we introduced a deep learning framework, ResNet-SEMultiRes, which combines a residual neural network backbone, a channel attention mechanism, and multi-resolution feature fusion for lung cancer discrimination on CT images. The architecture was trained and tested separately using two datasets: a public chest CT scans dataset of three cancer types and a private dataset of four cancer subtypes. According to the experimental results, the ResNet-SEMultiRes model achieves an average accuracy of 96.29% using the public dataset, which is higher than the 94.06% accuracy achieved by the ResNet-101 model. It also achieves an accuracy of 97.46% on the private dataset, outperforming the baseline by 1.97 percentage points. The above results demonstrate the robustness and generalization of performance on different datasets. The proposed ResNet-SEMultiRes model is effective for CT-based lung cancer classification and demonstrates better generalization ability across different datasets.