Enhancing Lung Cancer Diagnosis: Edge-TransUNet for Segmentation and Lightweight-MLCAN for Classification

As lung tumor is the leading contributor to tumor-associated mortality, accurately and efficiently identifying malignant lung nodules is paramount for efficient treatment planning and ameliorating survival outcomes. Existing diagnostic methods using Computer Tomography scans are time-consuming, prone to error, and rely heavily on expert radiologists, which leads to inconsistent outcomes. To address these challenges, this study proposes a novel Edge-based TransUNet for segmentation and a Lightweight Multi-level cascaded attention network for nodule classification, thereby enhancing the accuracy of diagnostic lung tumor diagnosis. The Edge-TransUNet incorporates a pyramid edge extraction module in the encoder region to derive promising edge information for delineating the boundaries of the lung nodules efficiently. For classification, the proposed classification approach employs MobileNetV2 as the primary feature extraction backbone and introduces a novel cascaded framework for accurate lung nodule classification. This novel cascaded framework aggregates the strengths and advantages of three modules, including a multi-level atrous convolution module for multi-level feature extraction, a second-order pooling module for gathering statistical data at the second order, and a multi-attention module for identifying the most representative features, which facilitates efficient lung nodule classification. Employing different lung computer tomography datasets, the Edge-TransUNet achieves a 0.981 dice coefficient, while the Lightweight-multi-level cascaded attention network accuracy of 98.5%, and precision of 98.3%, which outperforms other existing methodologies. Overall, this research offers a comprehensive and efficient solution for timely and reliable lung cancer diagnosis in medical applications.